Preliminary results of using an Intelligent Agent and Case Based
Reasoning for Quadratic Portfolio Optimization
Edward Falconer, Abel Usoro, Mark Stansfield, Brian Lees
There have been a number of intelligent agent systems
that have assisted portfolio management decisions.
The majority of these systems have focused on analysis
methods, such as fundamental analysis, technical
analysis or analysis of trader behavior. According to
Clarke et al, because of the efficient markets
hypothesis, portfolio managers should limit the time
that they spend on analysis of this type and instead
concentrate on other aspects of portfolio management,
such as diversification, asset allocation or cost
management. This study investigates the application of
intelligent agents and case based reasoning to strategic
asset allocation.
Traditionally, strategic asset
allocation is calculated using tools for quadratic
optimization and determining an investor’s utility
towards risk.
Associated with strategic asset
allocation are many complexities, such as the
distributed nature of problems that make a solution
difficult to implement. This paper outlines results of an
evaluation of whether intelligent agent and case based
reasoning could overcome these complexities. From a
review of literature there does not appear to have been
a study that has considered this problem in the way
proposed in this paper.
Keywords: Intelligent agents, quadratic optimization,
portfolio management, asset allocation, case based
reasoning.
1.
Introduction
The efficient markets hypothesis has led academics and
practitioners to divide investment management approaches
into two primary types: i) active investment management,
where an investment manager uses available information and
forecasting techniques to seek better performance than a
diversified portfolio; and ii) passive investment
management, where an investment manager has minimal
expectations about the returns that can be achieved and relies
instead on diversification to match the performance of a
market index (Clarke et al., 2001; Fama, 1991). Although
each of these approaches has benefits, there has been strong
evidence to support the claim that the profitability of active
investing is minimal, especially after transaction costs, and
therefore passive investing is more profitable (Clarke et al.,
2001; Bernstein, 2000; Bernstein, 2002). One of the
implications of choosing a passive investment approach is
that portfolio managers must concentrate on other strategies
for improving the return from their portfolio (Clarke et al.,
2001).
Examples of these strategies include cost
management, diversification and asset allocation strategies.
This research concentrates on asset allocation.
In general, asset allocation is a process whereby an overall
sum of cash is separated into different portions. Each
portion is then allocated to a different category of asset
class. According to Sharpe (1992, 2001) each asset class
should represent a capitalisation-weighted portfolio of
securities in order to mimic the return variation created by
different weights of asset class in the returns of the portfolio
under evaluation. There exist many different approaches to
asset allocation. These variations include strategic, tactical
and dynamic asset allocation, amongst others.
This
research concentrates on strategic asset allocation. The
optimal strategic asset allocation can be determined in one
of two ways: by maximizing return for a given risk level or
alternatively, minimizing risk for a particular return
objective (Nemati and Iyer, 1999; Kaplan, 1998; Todd and
Markowitz, 2000).
A strategic asset allocation decision is made up of a
quadratic component and a utility component (Nemati and
Iyer, 1999). The quadratic component is concerned with
constructing the portfolio under quadratic constraints whilst
the utility component is concerned with determining the
investor’s attitude towards risk. Moreno et al., (2005)
described Markowitz’s mean variance approach as the
classical model for risk forecasting and the preferred
approach for making strategic asset allocation decisions.
An earlier version of the work described in this paper was
presented at the International conference on intelligent
agents, web technologies and internet commerce – IAWTIC
– in Vienna Austria (20061).
This research considers the application of a combined
intelligent agent and case based reasoning (CBR) approach
to the strategic asset allocation problem. Broadly speaking
literature on intelligent agent and case based reasoning in
finance has focused on three areas: i) portfolio monitoring,
ii) stock selection, and iii) behavioural finance. Portfolio
monitoring is defined as the ongoing, continuous, daily
provision of an up-to-date financial picture of an existing
portfolio (Decker and Sycara, 1997). Stock selection is
defined as an analysis technique which uses valuation and
forecasting techniques to identify mis-priced assets (Davis
and Liu, 2002). Behavioural finance is defined as the
computational study of economies modelled as evolving
systems of autonomous interacting agents (Tetsfatsion,
2001). All of the artificial intelligent studies that have been
found have focused on analysis methods (such as
fundamental analysis, technical analysis or analysis of
1
The conference proceedings are currently in print
therefore page numbers cannot be provided. Expected
publication date is January 2006.
1
trader behaviour), which are aimed at identifying investment
opportunities with above average returns. The determination
of an investor’s utility or support of quadratic programming
does not appear to have been considered. This research
therefore assesses the effectiveness of intelligent agents and
CBR in dealing with the complexities associated with this
problem. This paper concentrates on the quadratic program
calculations under mean variance efficiency. Utility theory
will be considered in a subsequent paper.
The paper reports preliminary results on the effectiveness of
a prototype CBR and agent based system towards the
quadratic programming problem. The rest of this paper
provides: (i) an introduction to strategic asset allocation; (ii)
a background to the research problem and an overview of the
data preparation steps; (iii) an overview of agent and CBR
characteristics; (iii) an outline of the agent architecture; (iv)
an application of case based reasoning to the problem and
preliminary results; (v) conclusions and future directions.
2.
Strategic Asset Allocation
Strategic asset allocation refers to a situation where a
portfolio is constructed using mean-variance optimisers and
is considered a long-term investment strategy which is
modified at regular intervals. Tactical asset allocation is the
process of diverging from the strategic asset allocation when
an investor’s short-term forecasts deviate from the long-term
forecasts used to formulate the strategic allocation. Dynamic
asset allocation refers to strategies that continually adjust a
portfolio’s allocation in response to changing market
conditions.
The most compelling evidence for the use of risk-based
strategic asset allocation is provided by Brinson, Hood and
Beehower (referred to as BHB). According to Bitters (1997:
page 56) “the contribution they made was revolutionary”.
BHB questioned the importance of asset allocation and
compared the returns that are achieved by an investment
manager from factors such as: market timing, stock selection
and asset allocation policy. BHB questioned an investment
manager’s abilities to choose stocks, and time deals, against
their ability to allocate assets through an asset allocation
policy. BHB compared those adopting a strategic asset
allocation2 approach with those adopting a dynamic or
tactical asset allocation approach by measuring the portfolio
return achieved by introducing active analysis after making a
risk-based strategic portfolio choice. According to Bitters
(1997), BHB observed that:
(a) Investment policy return explained on average 93.6
percent of the total variance in actual plan return.
(b) Returns due to policy and market timing resulted in
average variance explained by 95.3 percent.
(c) Returns achieved by policy and security selection
returned in an average variance explained by 97.8
percent of the overall return.
BHB showed that on its own, risk-based strategic asset
allocation produced an average return of 93.6% and when
BHB actually used the term “asset allocation policy” which
was in vogue until Brennan et al (1997) coined the term
“strategic asset allocation”.
2
combined with non-strategic approach, the increase in
return was insignificant (4.2% at the most, ie the third case
above). BHB also discovered that investment managers
opting for an active approach, that is those using market
timing or stock selection techniques and tactical asset
allocation decisions, resulted in a 1.10 percent reduction in
the annual overall plan return (Bitters, 1997). In addition to
this, they discovered that involving active management
actually increased the risk associated with a portfolio and
produced less returns than portfolios constructed purely
through strategic asset allocation (Bitters, 1997). The
works of BHB highlight one of the major questions facing
the investment community, namely whether or not the
abilities of an investment manager can add any additional
profit to a portfolio. BHB concluded that investors are
better-off if they adopt a strategic asset allocation and
ignoring the alternatives. Bitters even went as far as to say
that investors are actually disadvantaging themselves by
opting for anything else. BHB provided compelling
evidence for the claim that investors should therefore
choose the passive approach and use strategic asset
allocation.
Unfortunately according to Rekenthaler (2000) the BHB
article was widely misquoted and led to the belief that asset
allocation was the answer the investment community had
been looking for. However, a preoccupation with asset
allocation led to a pre-occupation with asset allocation
methods, often to the detriment of the investor. According
to Rekenthaler (2000), the debate raged on until Ibbotson
and Kaplan (2000) published a paper that suggested the
returns could be 40, 90 or 100 percent of performance
depending on how the data was interpreted. Current
opinion is that strategic asset allocation remains a central
element in the process Rekenthaler (2000). However,
attempts at determining how useful are thought to be
fruitless (Rekenthaler, 2000). The view of many is that
100% of the return achieved on a portfolio is through asset
allocation (Jankhe, 1999). The evidence therefore remains
stacked in favour of the strategic approach.
3.
Mean Variance Efficiency
Portfolio selection (also known as Mean Variance
Efficiency) considers the problem of efficiency and the
identification of what Markowitz (1992) defines efficient
set analysis. Portfolio selection and portfolio efficiency are
about variance, return and risk. A portfolio is considered
efficient if the mix of assets between bonds and stock
investments, and the variance in return from those
investments, meet the risk profile of an individual investor.
Thereby a portfolio is considered inefficient if another
portfolio can be constructed with a higher expected return
and the same variance, or with lower variance and the same
expected rate of return (Markowitz, 1992). In 1990
Markowitz was awarded a share of the Nobel Prize in
economics for his contribution.
According to Markowitz the investor could be seeking more
risk or they could be seeking less of it. This depended on
their expected rate of return. The problem therefore
became a quadratic equation with the amount invested in
each security determined by the risk preferences of the
2
investor and the risk and return associated with the
securities. Investing in securities with different rates of risk
and return balances out the risks involved if the investor
chose only one security for their portfolio. The belief is that
the return on a portfolio would be protected from losses if an
investor chose securities that have different variances and
covariances. Markowitz proved that this was the case under
quadratic constraints.
The aim of quadratic programming in the portfolio selection
problem is to calculate the amount that should be invested in
different securities based on the covariance and expected
returns related to those securities. The quadratic program
uses a value representing the investor’s risk aversion as an
input variable and adjusts the amount that is invested in each
security based on either the expected returns or the
covariance of return, within a portfolio. Markowitz (1952,
1959, 1992) defined the following objective function to
determine how to maximise returns or minimise risk subject
to an individual’s risk preferences.
Min(max)imise Σ µi xi – p Σ Σ σij xi xj
Subject to:
Σ xi = 1 &
xi >= 0 for all i.
In the function, p represents the investor’s aversion to risk,
µi represents the return on the ith security, xi represents the
amount invested in the i’th security and σij refers to the
covariance of the ith and jth security. The simplest way of
implementing a solution under quadratic and linear
constraints is to substitute the investor’s risk aversion for a
value representing the minimum portfolio return (Fylstra,
2005). This minimum return value then replaces the letter p
in the equation and allows the portfolio to be determined
based on return and risk. By taking this approach the
quadratic programming problem is about which asset classes
have been chosen and the minimum portfolio return.
4.
Data Preparation
The expected return and covariances used within the
quadratic calculation are determined using historical index
pricing information. In contrast to the standard approach to
portfolio selection, the securities, in this case, are in the form
of global asset prices. The global approach differs slightly to
the standard asset allocation approach as the focus is placed
upon prices that reflect all equities within a specified global
market. In contrast, asset allocation approaches usually
focus on much broader asset allocation classes, such as real
estate, stocks, bonds and cash. Despite the differences, there
are many advantages to adopting a global approach. These
advantages include global asset coverage and the availability
of daily price updates. This decision is also supported by the
globalisation of markets discussed in literature and the
importance of global diversification (French, 1991; Lummer
et al., 1994a; Lummer and Riepe, 1994b; Kaplan, 1998;
Dempster et al., 2003). Gerber (1994) and Black and
Litterman (1992) provide compelling evidence for the use
of global asset classes. In particular, they discuss the
benefit that is gained from broad diversification across
capital markets and the high degree of correlation in return
that appears at times within these markets. Lummer and
Riepe (1994a; 1994b) and Kaplan (1998), also discuss how
global asset allocation provides the opportunity to apply
Mean Variance Efficiency more effectively because of the
general stability of global markets.
The historical global asset prices were obtained from
Morgan Stanley Corporate International (MSCI©).
According to the MSCI website, ‘close to 2000
organisations worldwide currently use the MSCI
international benchmark indices…MSCI estimate that over
USD 3 trillion are currently benchmarked to these indices
on a worldwide basis’. The following global equity asset
classes were used for the calculation: UK, Japan, Australia,
Thailand, Colombia, Brazil, USA, Canada and Germany.
This selection of asset classes provides broad coverage
across developing and established markets. Although the
assets are limited to those specified, the model could easily
be extended to include additional asset classes, cash assets
or bond assets, with minimal effort.
Under portfolio selection theory, return and risk is
determined by obtaining the following values for each of
the chosen asset classes (Markowitz, 1992):
Ci – Pi + Ii
Ci
The letter C represents the closing price from the current
year, the letter P represents the closing price from the
previous year and the letter I represents the income from the
index. The lower case i represents the year. Each mean
return is then represented in terms of a percentage returned
from a specific year. In this study, the mean variance
calculation was carried out using ten years worth of asset
prices. Annual returns and mean returns are used to
determine the extent to which an individual asset class
deviates from the average asset class price. In addition to
this, the variances from the average return are also used in
the calculation.
The following equation is used to
determine the deviation of the annual return:
Deviation from the Average Return = (ri) – (r’i)
The letter r is the return variable (annual return), r’ is the
average return and lower case i represents the year. The
standard deviation of return is then determined from the
overall portfolio variance. Table 1 provides an example of
the annual return, average return and standard deviation of
return, of the UK market over a ten year period (priced in
US dollars).
Name: UK
Price: USD
3
Date
Annual Price
Dec 1993
Dec 1994
Dec 1995
Dec 1996
Dec 1997
Dec 1998
Dec 1999
Dec 2000
Dec 2001
Dec 2002
Dec 2003
Dec 2004
641.457
611.077
716.422
883.481
1,052.09
1,208.25
1,325.94
1,146.25
962.025
791.076
1,006.08
1,162.44
Average (Mean)
Standard deviation =
Annual
Return
Deviations
Squared deviations
-0.049715502
0.147043223
0.189091786
0.160264197
0.129239303
0.088759186
-0.156755832
-0.19150126
-0.21609681
0.213705457
0.134511542
0.040776844
0.154781582
0.090492347
-0.106266379
-0.148314941
-0.119487353
-0.088462458
-0.047982342
0.197532677
0.232278105
0.256873654
-0.172928613
-0.093734697
0
0.008188865
0.011292543
0.021997322
0.014277227
0.007825607
0.002302305
0.039019158
0.053953118
0.065984074
0.029904305
0.008786193
0.023957338
Table 1:Computation of Standard Deviations, Mean Returns and Variances, using broad asset classes (Calculations based on Markowitz's theory of meanvariance efficiency)
Covariance and correlation coefficient were then calculated
across each of the global market asset classes (these
correlation coefficient matrix and expected returns have been
detailed in Appendix 1). Groups of assets were then chosen
and the coefficients from the countries were loaded into a
quadratic program (QP) which performed optimisation
calculations. The QP was then configured to maximise
return or minimise risk using a variety of pre-determined
constraints. The quadratic program, as previously explained,
uses the variables that it is given to calculate how much must
be invested in each asset class to achieve a certain level of
return for a certain level of risk.
A quadratic program named MOSEK© was used to perform
the calculations. In order to choose assets a front end
application was developed. The user could also choose a
specified minimum portfolio return. The data was then fed
into the quadratic program for the calculations to be made.
The covariance and mean return values for each of the
chosen assets are collected from a variety of flat CSV
(comma separated variables) files. Prior to passing the
information to the quadratic program the data was formatted
in a way that could be interpreted by the quadratic program.
The outputs of the quadratic program are suggested
investment weightings which would maximise return subject
to the constraints applied. The prototype software agent and
CBR are written in VB.NET and use a class library (via a
VB.NET API link to MOSEK©) to link to the quadratic
program.
5.
Agent & CBR Characteristics
Strategic asset allocation is characterised by a variety of
complexities that have to be considered in the design of a
decision support system (Nemati and Iyer, 1999). These
complexities are related to the problem of matching
investors with portfolios and the overall application of the
strategic asset allocation theory. The complexities are listed
below.
(i)
The strategic asset allocation task is distributed over
various parts; this means that different factors
contribute to the matching of an investor with a
portfolio.
(ii) Some tasks require input from other tasks to arrive at
a conclusion; thereby necessitating the need to share
data and control the flow of information between
tasks.
(iii) There is a chance that the same set of data inputs
could be used multiple times; this means there is a
potential for task duplication.
(iv) There is not one solution that can be applied to every
set of circumstances; this means that each portfolio
or investor could be either unique or at best similar
but often not exactly the same as those encountered
previously.
(v) In order to match up an investor with a portfolio it is
necessary to have input from a user; and this user
could require a different level of assistance
depending on their preferences.
In the light of these observations a review was conducted of
available technologies and intelligent agent and case based
reasoning were identified as displaying the types of
attribute that could deal with the complicated nature of the
strategic asset allocation decision.
Agent and CRB
technologies have been used to successfully deal with the
complexities of portfolio management (Davis and Luo,
2001; Davis et al., 2000; Decker and Sycara, 1997).
5.1
Intelligent Agents
4
Intelligent agents are software artefacts that have a number
of characteristics that distinguish them from other
technologies. In particular, an agent can be given a goal or
task which it can complete autonomously (Turban and
Aronson, 2001). Agents may also carry out tasks or goals in
parallel with the existing computer operations and ask for
advice from humans if they become “stuck” in a problem
(Wooldridge, 2002).
Agents can act proactively and
communicate with human users, providing a personalised
service to some users and communicating with other agents
in a multi-agent environment (Turban and Aronson, 2001).
incorporate this autonomy by taking the collection of
knowledge and forming the suitably structuring the data.
The overall aim is usually for the intelligent agent system to
support the problem solving activities of the human expert.
Turban and Aronson (2001) also discuss how over time it
may be necessary to maintain the knowledge base, by
identifying and correcting any inconsistencies, and by
filling in any perceived gaps through the acquisition and
inclusion of additional knowledge. The ability to develop
knowledge and interact with a user is also required within
the solution.
Wooldridge (2002) discuss how intelligent agents can
interact with, and react to, a user’s preferences. This means
that software can be designed to recognise a user’s actions
and modify the user/agent interaction to ensure the level of
support is appropriate to that user. Decker and Sycara
(1997) also discuss how intelligent agents can incorporate a
separate control component to regulate the flow of
information between distributed tasks.
This control
component can organise the flow of information that passes
between the various parts of the problem and ensures that a
conclusion is reached. This ability to control distributed
tasks is a key attribute of the anticipated solution.
Intelligent agents therefore demonstrate various abilities
that suggest they may be appropriate for resolving this
problem. In particular the ability of intelligent agents to
interact with human experts, organise data flow, construct
knowledge bases and learn from experience are perceived
to be important for overcoming the complexities associated
with matching portfolios with investors.
In order to build the proposed solution it is necessary to
acquire knowledge from a variety of knowledge sources:
human expertise and recorded knowledge of previous
successful problem solutions. This means the system must
have an element of autonomy. Intelligent agents can be
5.2
Case Based Reasoning
Case-based reasoning (CBR) is a methodology that can
allow software to learn from experience. Case based
reasoning subdivides a problem into a series of tasks and
then combines the results of each task to form a case. Each
case is then compared with previous cases to determine
whether new experiences can be learned. Fig 1 illustrates a
general CBR cycle (Althoff, 2001) the figure illustrates the
four tasks that exist as a part of all CBR programs.
New Problem Situation
Retrieve
Similar
Cases
Reuse case
information
Case Base
Retain
experience
Revise
proposed
solution
Proposed Solution
Fig 1: General CBR Program Cycle
Fig 1 identifies four tasks: Retrieve task – to retrieve similar
cases; reuse task – to reuse information and knowledge from
previous cases to solve a problem; revise task – where the
problem is revised in line with previous experiences; retain
task – where experience is retained that could be used for
future cases. Using this general cycle a software program
can learn from its experiences.
CBR was found to offer the types of capabilities that may
make it appropriate for dealing with the complexities of the
strategic asset allocation problem. For instance Althoff
(2001) discusses how CBR systems gather case data from
distributed tasks and join the outputs of these tasks together
to formulate a combined state. This combined state then
contributes to any future development. In this way a
5
distributed problem can be resolved, with each distributed
solution contributing to all future solutions. Leake (1996)
also discusses how CBR systems can be used to subdivide
each solution into different parts and then pass requests for
analysis onto other components. By initiating this task
division CBR software can perform distributed tasks at the
same time as other tasks are performed and then combine the
parts to arrive at a more complete solution.
Althoff (2001) identifies that CBR systems evaluate each set
of results to determine whether the current case is new, or
whether it matches a previous case. The CBR system then
adds new cases and modifications to existing cases to the
case base for the future benefit of the software. Each case is
made up of a set of responses that are unique to each
problem and solution. By comparing current information
and approaches with new sets of input CBR programs are
able to identify similarities and differences with each new set
of inputs. In so doing the CBR system therefore learns from
its experiences.
CBR was therefore found to offer additional capabilities
that would strengthen the intelligent agent approach. It is
the ability of CBR to organise distributed tasks, subdivide
solutions into categories, historically evaluate solutions and
learn from experience that mean it may be applicable for
dealing with the complexities associated with strategic asset
allocation.
6.
Intelligent Agent Architecture
The purpose of the agent architecture is to define how the
tasks are separated and organised. The agent architecture is
designed to carry out distributed tasks separately then
combine the results to make a strategic asset allocation
decision for the user. The agent architecture is shown in
Fig 2. There are five parts to the agent: (i) investor profiler;
(ii) asset selector; (iii) control component; (iv) case based
reasoning component; and (v) maths calculator. Each
component is outlined below.
User
Question
Database
Investor
Profiler
Asset
Selector
Asset
Pricing
Files
Control
Component
QP
(Mosek)
Maths
Calculator
CBR
Program
Case
Base
Fig 2: Prototype Architecture for Asset Allocation
Investor Profiler: The investor profiler retrieves questions
from a questionnaire class linked to a database. The
database stores a series of questions which it asks the user.
Each question also has a set of responses. The investor
profiler asks the user the questions and presents some
possible answers. The investor then chooses one of the
answers from those presented. The investor profiler then
writes these results to the database and sends the investor’s
risk profile to the control component.
Asset Selector: The asset selector component is tasked with
choosing assets that match the profile of the investor. The
choice of asset is evaluated based on the information passed
by the control component which initially comes from the
investor profiler. The asset selector retrieves input in the
form of asset prices from asset pricing files. It passes the
relevant information to the control component. The asset
selector component also communicates the results to the
user.
Control Component: The control component is tasked with
controlling the flow of information between the agent’s
various parts. The control component is used primarily to
control the flow of information within the agent
environment. The control component is tasked with passing
information for processing and passing processed
information around the agent environment.
The Maths Calculator: The maths calculator component
interacts with the quadratic program and modifies the
6
constraints, number of variables and other inputs, to allow
the quadratic program to make its calculations. The maths
calculator receives requests from the control component
about possible inputs to the quadratic program. The maths
calculator determines the inputs and reads the solution. It
plays the part of a driver program for information being
passed to the quadratic program (QP) program. Mosek was
used to perform the quadratic calculations.
The Case Based Reasoning component: The CBR
component interacts with the control component and tracks
information received from the different parts of the agent
architecture. It uses inferences around the inputs that have
been received previously to determine whether information
can be used again, or whether parts of a previous problem
can be used again. The CBR sends its results to the control
component. A unique reference method is used to ensure the
appropriate results are sent to the appropriate component.
The agent program is designed to support the activities
associated with the asset allocation task. The remainder of
this paper outlines preliminary results for the maths
calculator and provides an overview of the work undertaken
into the application of CBR to the problem. The additional
components will be covered in a separate paper.
7.
Preliminary Results
There are two sets of results to report, findings related to: (i)
the math calculator component and its interface with the
quadratic program; (ii) the application of CBR to quadratic
programming problem. Although the results are related,
they are considered separately as they were found to each
pose a different set of challenges, and were complex to
explain when combined together.
7.1
The Quadratic
Calculator
Program
and
Math
The QP used a quadratic function to determine the optimum
investment amount to place in each asset class. The
following objective function was used to maximise return or
minimise risk (from an article by Fylstra (2005).
maxmise
cTx - xTQox
subject to
Ax >=b
Σx = 1
xi >= 0 for all i.
o
Q is a matrix containing covariances for the assets held
within the portfolio. cT is a vector containing returns on
each of the assets held within the portfolio. A is also a
vector containing the returns on each of the assets held
within the portfolio. b is a value that represents the
minimum portfolio return. x is the amount to be invested in
the ith asset class within the portfolio. Within the function,
xTQox represents the risk associated with the portfolio and
cTx is the return on the portfolio. The values associated
with the portfolio are then adjusted based on the minimum
portfolio return and the asset selector’s choice of asset
class. Within the equation the constraints (Ax >=b; Σx = 1;
xi >= 0 for all i) remain the same, with the values
associated with the Qo, cT and A, adjusting in line with the
input of the problem variables. The maths calculator has a
front end application which allows a user to choose asset
classes, choose a minimum portfolio return, view
covariances and mean returns, and view the optimised
result. The program is illustrated in Figure 3.
The math calculator retrieves the values for the matrices Qo,
cT and A from a set of flat files. A separate flat file exists
for each of the mean returns (cT and A) and the covariances
(Qo). The data tables are detailed in appendix I. In order to
make the calculations, the user first chooses two or more
asset classes from those available and selects a minimum
portfolio return. Once the user has chosen the asset classes
and minimum portfolio return, they press the optimise
button and initiate the quadratic calculations. On selecting
the optimise button the math calculator first retrieves from
the asset selector which obtains the data from the flat files.
The maths calculator then formats the data for the quadratic
program. This formatting stage involves locating and
retrieving the mean returns and covariances associated with
each of the chosen assets and leaving the irrelevant data.
The math calculator then interfaces with the quadratic
program and requests the quadratic results.
The
calculations that are then returned are displayed in the
investment amount textbox. As well as this the math
calculator also displays the mean returns and covariances in
the respective textboxes.
7
Covariances
(Qo)
Mean Returns
(cT, A)
Minimum
Return
(b)
Optimal
Amount
(xi)
Figure 3: Maths Calculator with Example Quadratic Variables and Results
When interfacing with the quadratic program the math
calculator uses the values shown in Table 2 to initiate and
format the data. These values are determined from either the
data held within the matrices or the chosen asset classes.
Each value serves a dual purpose: (i) to enable the maths
calculator to input the appropriate data to allow the MOSEK
tool to calculate the results; and (ii) to allow the CBR
program to determine inferences about the data. The two
constraints - namely Ax >=b and Σx = 1 - remain the same at
all times with b changing in line with the investor’s
minimum portfolio return. The size of the three matrices and
vectors (Qo, cT and A) increase in line with the number of
variables included in the portfolio.
Value
Number of
variables
Number of
variables in the
matrix A
Number of
variables in the
matrix Qo
Values within the
matrix A
Values for the
matrix cT
Values within the
matrix b
Values within the
matrix Qo
Description
The number of asset classes chosen by
the asset selector
The same as the number of asset classes
chosen by the asset selector
Calculated using the following equation
(where N is the number of asset classes
chosen by the user):
N(N + 1)
2
The expected return for each asset class
The same as the expected return for each
asset class
The minimum portfolio return
The covariance for each value Ni, Nj
Table 2: Variables used by the MOSEK tool and determined by the maths
calculator
Despite the initial simplicity of the quadratic program in
calculating the optimum results, from close inspection, it
was apparent that at times it was not possible to produce
results that were useful. An extensive evaluation was
therefore conducted of the quadratic program and the
outputs that were produced under various constraints. This
section introduces results of an extensive evaluation that
was conducted of the quadratic program and the outputs
that were produced under various constraints.
This
evaluation involved evaluating 675 variations of portfolio
choice. The range of choices was created by adjusting the
minimum portfolio return (which ranged from 0.01 to 0.15)
for the full range of 45 different asset allocation variations
(with 9 asset classes). This evaluation was useful as it
outlined the types of limitations that could be encountered
by using a quadratic program. In particular from this
evaluation it was revealed that at times it was not possible
to produce results that were useful. This problem is
discussed by Fylstra (2005) who indicates that analysts take
great care to ensure that their portfolio models remain
calculable under quadratic or linear constraints. These
problems are encountered when an invalid choice of asset
class and minimum portfolio return combination are
chosen. The following four categories identify the types of
result that were returned.
1.
In the first category the optimal result is split based on
the variables and constraints as they have been defined.
The program therefore supplies optimal amounts that
provide beneficial advantage and are sensible from an
asset allocation perspective. Table 3 provides an
example of a set of quadratic results that meet this
condition.
Country
USA
Canada
Germany
UK
Total
Variable X
X0
X1
X2
X3
Optimum investment amount
0.289805523
0.306632754
0.189994446
0.213567277
1.000000000
Table 3: Optimal results for the quadratic code and sample
inputs
8
2.
In the second category the results are infeasible. This
arises because the variables make the equation nonlinear and therefore not solvable by a quadratic
equation. The example in table 3 shows the results
returned for infeasible solution. Specifically, the total
shows as 4 and thus fails the first test of feasibility, ie
the constraint that summation of x’s should be equal to
1 (i.e. Σx=1).
Country
Japan
Australia
Thailand
Columbia
Total
Variable X
X0
X1
X2
X3
Optimum investment amount
1
1
1
1
4
Table 4: Results obtained for a Infeasible quadratic equation
3.
In the third category although the results have passed
the first test of feasibility (Σx=1), they are so skewed
towards one or two asset classes that the costs of
investing in the remaining asset classes outweigh the
benefits. An example of this is shown in table 5. In this
example the optimal result would involve investing
0.99999998504% in Germany and UK, but investing
only 0.00000000576% in the USA and Canada In this
instance In . This although mathematically accurate
(Σx=1) is not beneficial from an investing perspective
because the cost would outweigh the benefit.
Country
USA
Canada
Germany
UK
Total
Variable X
X0
X1
X2
X3
Optimum investment amount
0.00000000472
0.00000000104
0.06442543178
0.93557455326
1.00000000000
Table 5: Results obtained from a combination of asset
classes
4.
In the fourth scenario the results break the constraints
that the sum of x must equal 1.
Table 6 shows an
example of this scenario. This was found to occur
whenever a combination of asset classes and minimum
portfolio returns caused the equation to become nonlinear in some places but linear in others.
Country
USA
Canada
Germany
UK
Total
Variable X
X0
X1
X2
X3
Optimum investment amount
0.1284890814
0.1415283229
0.0682738176
0.0803402387
0.4186314686
Table 6: Results obtained where the portfolio is not fully
invested
It was found from the evaluation that each quadratic result
fell into one of these four categories. If category two
occurred the user was asked to change their choice. If
category three occurred the amount was proportionally
adjusted so that the amount was equally split. If category
four occurred the user was asked to change their choice. The
CBR program is also being incorporated to resolve these
problems by allowing the program to evaluate the user’s
choice and identify suitable alternatives. These alternatives
are aimed at improving the solution.
7.2.
Case Based Reasoning Search and Retrieve
Method
This section outlines the development of the case-basedreasoning solution and the identification of the case
features. The primary purpose of the CBR component is to
assist with the problem of matching investor’s with
portfolios and ensuring that the agent system makes
efficient use of resources and time. The CBR program
operates by first having the quadratic program and investor
profiler carry out their tasks then providing the problem and
solution variables to the CBR component. The CBR
program then evaluates and monitors cases as new
problems and solutions are introduced. The intension will
be that as time passes the quadratic program, investor
profiler and stock selector program will be used less and
will be replaced by the solution that is provided by the CBR
program. In general the CBR tool carries out the following
steps:
1. Searches from its memory to find portfolios that
have been previously been chosen for investors with
specific responses and makes an assessment of
similarity.
2. Infers an answer from the most similar matches.
3. Adjusts the portfolio to investor matching solution
for changes in circumstances by inferring relationships
between multiple sets of the historical data.
4. Modifies solutions from step three and records
them for future use.
The search and retrieve stages are completed in the first and
second steps and will use the nearest neighbour retrieval
method. The nearest neighbour retrieval method identifies
possible matches by defining values that can be used to
index each case. The indexes are then weighted in terms of
importance. Each weight is used to calculate how far away
a previous case is from the current case. Using the nearest
neighbour retrieval method there are three states that the
CBR program can find itself in when evaluating
alternatives: (i) an exact match that is obtained by
identifying that the same set of input values within the
problem have been used before; (ii) no match that is
obtained if the current case does not match any of the
previous cases; or (iii) a partial match that is obtained if the
current case is similar to previous cases, but not exactly the
same. The distance between previous cases and a current
case is proportional to the similarity between the features of
each case. Previous cases given a partial match can be far
away or close to the current case. However, the closer a
previous case is to the current case the greater the chance
that it will be considered as a possible match.
Rule Description
CountryName
NumberOfCountries
GlobalInfluence
AverageVolume
Exact
Match
10
10
5
5
Partial
Match
0
5
2
2
No
Match
0
0
0
0
9
LengthEstablished
EmergingRating
Covariances
MinimumPortfolioReturn
MeanReturn
Mosek Result
RiskAttitude
RiskPerception
KnowledgeAndExperience
OptimalRiskLevel
Complexity
InvestmentGoal
InvestmentObjective
InvestmentTimeHorizon
InvestmentAmount
EstimatedYearReturn
InvestorExpectedReturn
Total
5
5
10
10
10
10
10
10
5
5
10
10
10
10
10
5
5
180
2
2
5
5
5
5
5
5
2
2
7
5
5
7
7
2
2
87
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Table 7: Case Features as they apply to the quadratic problem and
investor profiler
The purpose of the CBR program will be to categorise the
portfolios chosen for specific investors so that previous cases
can be compared as possible alternatives. Each case has a
set of features that are used to measure how similar a
previous case is to a current case. The features that relate to
this particular problem have been identified and are shown in
table 7. Each feature has been allocated a weight based on
the importance of that feature to the problem. Therefore for
instance the country name and number of countries are
considered more important than global influence, average
volume etc. This is apparent because country name and
number of countries have a score of 10 whilst global
influence and average volume have a score of 5. Each
feature either relates to an aspect about either the portfolio or
the investor.
The features are used to determine how close a previous case
is to a current case.
The purpose is to categorise the
portfolios based on the case information that is attained. The
portfolios will be categorised with the score relating to each
portfolio and investor combination. This score is then used
to determine how close this case may be to previous cases.
The portfolios will be categorised by the level of risk they
contain. The values will be interpreted by considering the
potential risks associated with the individual portfolio as it
compares with other portfolios. An example of the matching
strategy being adopted is noted below.
Step 1: User chooses a portfolio chosen. During this step
the user is asked to choose asset classes from the checkbox
and selects a minimum portfolio return and selects the
optimise button. In this example the user has chosen a
portfolio containing the USA, UK, Germany and Canada
with a minimum portfolio return of 0.15. The results of this
calculation have already been shown in table 6. The user
also provides input to the program that allows the investor to
be characterised in terms of investment objectives, time
horizon and knowledge and experience of investing.
Step 2: Previous cases retrieved. During this stage all
previous cases are compared with the portfolios chosen in
step 1 for the type of investor identified by the user. This
retrieval is based on the values identified in table 7. In total
three portfolios have been found that have been chosen for
investors with similar characteristic. These cases are each
considered as a potential match for the current case.
The first portfolio contains USA, UK, Thailand and Canada
and has the same portfolio return as the current portfolio.
In investor has some similarities to the current investor
however the time horizon and risk profile are different. The
first portfolio/investor combination achieved a 58%
similarity and received the following scores:
CountryName
NumberOfCountries
GlobalInfluence
AverageVolume
LengthEstablished
EmergingRating
Covariances
MinimumPortfolioReturn
MeanReturn
MosekResult
RiskAttitude
RiskPerception
KnowledgeAndExperience
OptimalRiskLevel
Complexity
InvestmentGoal
InvestmentObjective
InvestmentTimeHorizon
InvestmentAmount
EstimatedYearReturn
InvestorExpectedReturn
Score
5
10
2
2
2
2
5
10
5
10
5
5
2
2
10
5
10
0
10
2
2
103
The second portfolio contains USA, UK, Canada and has
a different portfolio return, however, the investor has
different goals and investment amount. The second
portfolio/investor combination achieved a 60% similarity
and received the following scores:
CountryName
NumberOfCountries
GlobalInfluence
AverageVolume
LengthEstablished
EmergingRating
Covariances
MinimumPortfolioReturn
MeanReturn
MosekResult
RiskAttitude
RiskPerception
KnowledgeAndExperience
OptimalRiskLevel
Complexity
InvestmentGoal
InvestmentObjective
10
0
5
5
5
5
5
10
5
10
5
5
5
5
10
0
5
10
InvestmentTimeHorizon
InvestmentAmount
EstimatedYearReturn
InvestorExpectedReturn
Score
7
0
5
2
109
The third portfolio contains USA, UK, Canada, Germany
and has a different portfolio return. The investor however is
very similar to the current set of investor responses provided
by the investor. The third portfolio/investor combination
achieved an 87% similarity and received the following
scores:
CountryName
NumberOfCountries
GlobalInfluence
AverageVolume
LengthEstablished
EmergingRating
Covariances
MinimumPortfolioReturn
MeanReturn
MosekResult
RiskAttitude
RiskPerception
KnowledgeAndExperience
OptimalRiskLevel
Complexity
InvestmentGoal
InvestmentObjective
InvestmentTimeHorizon
InvestmentAmount
EstimatedYearReturn
InvestorExpectedReturn
Score
10
10
5
5
5
5
5
5
10
10
10
10
5
5
10
10
10
7
10
5
5
157
Step 3: Reuse case information. As described in diagram 1
this stage is completed based on the results of the previous
step. Of the three portfolio/investor combinations that have
been identified as possible matches the third portfolio most
accurately meets the set of criteria. In order for the current
problem to be identified as a possible alternative the
minimum-portfolio-return and the features relating to the
investor are all required to be very close to that of the current
case. All three portfolios are displayed as possible matches
to the user with the score displayed for each portfolio.
The final stages of general CBR cycles (identified in
diagram 1 as revise proposed solution and update case base
and retain experience) have not been considered in this paper
and will be considered at a later stage. One of the
advantages of adopting CBR is that it is possible to allow the
system to evolve and grow with time. This makes the
quadratic solution particularly interesting for determining
alternative solutions to a given problem. The distributed
nature of the problem means that there are an exponential
number of different combinations that could occur.
Although the laws of averages would suggest that the same
combination of results will be chosen over and over again,
this cannot be confirmed. By incorporating CBR it is
possible to cater for both the scenario where the situation
has or has not occurred before and allow the system to learn
from its encounters. This is especially beneficial due to the
number of calculations that must be made to produce a
complete data set. For instance, even with the numbers of
asset classes reduced to 9 (as is the case here) the volume of
calculations required for any portfolio optimiser to
complete a proper evaluation of these cases reaches into its
thousands. With the introduction of additional asset classes
the process could therefore become even more problematic.
8.
Conclusions and Future Direction
This paper outlines a decision-support system to assist with
strategic asset allocation decisions that is based on an
intelligent agent. The system uses case-based reasoning to
make portfolio optimisation calculations under quadratic
constraints. These calculations are performed as part of an
optimisation of the expected return and the minimisation of
risk within a portfolio of broad asset classes.
The results attained from an evaluation of the maths
calculator present sufficient evidence to support the
decision to use intelligent agents for the problem of
matching investors to portfolios. In addition to this the ease
with which the variables associated with the maths
calculator fit with the general approach to CBR also support
the decision to use CBR within the approach.
Combining the capabilities of intelligent agent and CBR
based systems therefore present the opportunity to support
strategic asset allocation decisions. By combining these
technologies the system could learn from its interactions
with the user and also its experiences when evaluating
problems.
The next piece of work to be undertaken will involve the
development of the additional agent components.
Specifically this refers to the investor profiler which will
consider utility theory in its determination of a final
solution. Work will also be undertaken on extending the
maths calculator so that it carries out conic optimisation
calculations as well as quadratic calculations.
References
Althoff, K (2001), “Case-Based Reasoning”, ICCBR-01 Workshop
On Case-Based Reasoning Authoring Support Tools, 31st
July.
Bernstein, W (2000) The Intelligent Asset Allocator: How to Build
Your Portfolio to Maximize Returns and Minimize Risk,
McGraw-Hill Education.
Bernstein, W (2002) The Four Pillars of Investing: Lessons for
Building a Winning Portfolio, McGraw-Hill Education.
Bitters, W.E (ed.) (1997) The New Science of Asset Allocation,
Glenlake Business Monographs.
Black, F; Litterman, R (1992) “Global Portfolio Optimization”,
Financial Analysts Journal, Sept, 28-43.
Brinson, G.P., Hood, L.R., Beebower, G.L., 1986, “Determinants
of portfolio performance.” Financial Analysts Journal.
July/August, pages 39-44.
11
Campbell, JY; Viceira, LM (2002) Strategic Asset Allocation:
Portfolio Choice for Long-Term Investors, Oxford University
Press.
Clarke, J; Mendelker, G; Jandik, T (2001) Expert Financial
Planning : Investment Strategies from Industry Leaders,
(Edited by Robert C Arffa), 2001, chapter 10, Wiley and Sons.
Davis, D; Luo, Y (2001) “Using KADS to Design a Multi-Agent
Framework for Stock Trading”, Multi-Conference Event Agent for E-Business on the Internet, Las Vegas, June 2001.
Davis, D; Luo, Y; Liu, K (2000) “A Multi-Agent Framework for
Stock Trading”, World Computing Conference, Beijing,
August.
Decker, KS; Sycara, K (1997) “Intelligent Adaptive Information
Agents”, Journal of intelligent Information Systems, Kluwer
Academic Publishers, volume 9, pp 239 – 260.
Dempster, Germano, Villaverde, (2003) “Global Asset Liability
Management”, British Actuarial Journal, Volume 9, Number
1, pp. 137-195(59).
Fabozzi, FJ; Markowitz, HM (2002) The Theory and Practice of
Investment Management (Frank J. Fabozzi Series), John Wiley
& Sons.
French, K (1991) “Investor Diversification and International Equity
Markets”, American Economic Review, MIT Press.
Fylstra, D (2005) “Introducing Convex and Conic Optimization for
the Quantitative Finance Professional”, Wilmott Magazine, pp.
18-22.
Gerber, G (1994) Equity Style Allocations: Timing Between Growth
and Value, in Global Asset Allocation: Techniques for
Optimising Portfolio Management, New York: John Wiley &
Sons.
Kaplan, P (1998) “Asset Allocation Models Using the Markowitz
Approach”,
http://www.misp.it/doc/materiali_doc/Curti1EN.pdf, accessed January 23rd January 2006.
Lummer, S; Riepe, MW (1994a) "Taming Your Optimisation: A
Guide through the pitfalls of Mean Variance Optimisation",
Global Asset: Techniques for Optimising Portfolio
Management, Edited by J.S. Lummer, M.W. Riepe, , John
Wiley & Sons.
Lummer, S; Riepe, MW (1994b) "The Role of Asset Allocation in
Portfolio Management", Global Asset Allocation: Techniques
for Optimising Portfolio Management, Edited by J.S.
Lummer, M.W. Riepe, John Wiley & Sons.
Luo, Y; Davis, D; K, Liu (2001) “Information and Knowledge
Exchange in a Multi-Agent System for Stock Trading”,
Appendix 1
Covariances and Mean Returns for 9
Classes (MSCI©)
Correlation Coefficients
USA
Canada
Germany
UK
Japan
Australia
Thailand
Colombia
Brazil
USA
0.03502
0.03547
0.04672
0.02897
0.04817
0.04817
0.17519
0.07985
0.07527
IEEE/IEC Enterprise Networking Applications and Services
Conference, (EntNet2001), Atlanta, USA, July.
Luo, Y; Davis, D; Liu, K (2002) “A multi-agent system
framework for decision making support in stock trading”,
The IEEE Network Magazine Special Issue on Enterprise
Networking Services, Volume 16, No 1, Jan/Feb.
Moreno, D; Marco, P; Olmeda, I (2005) “Risk forecasting models
and optimal portfolio selection”, Applied Economics, Volume
37, Number 11, 20, June, pp. 1267-1281(15), Routledge, part
of the Taylor & Francis Group.
Nemati, H; Iyer, LS (1999) “Intelligent decision support system
for asset allocation”, Americas Conference on information
Systems.
Paolucci, M; Niu, Z; Sycara, K; Domashnev, C; Owens, SR
(2002) Van Velsen. M., “Matchmaking to Support Intelligent
Agents for Portfolio Management”, In Proceedings of AAAI
(Demo Session), 2002.
Rekenthaler, J (2000) “A Different Twist on Asset Allocation”,
Journal of Financial Planning, January Issue.
Seo, Y. W; Giampapa, J. A; Sycara, K (2002) "Text classification
for intelligent agent portfolio management," First
International Conference on Autonomous Agents and MultiAgent System.
Sharpe, W (1992) “Asset Allocation: Management Style and
Performance Measurement”, The Journal of Portfolio
Management, Winter 1992.
Sharpe, W (2001) Budgeting and Monitoring the Risk of Defined
Benefit
Pension
Funds,
http://www.stanford.edu/~wfsharpe/art/q2001/q2001.htm,
2001. [accessed 10 October 2006]
Sharpe,
W
(2004)
http://www.stanford.edu/~wfsharpe/art/princeton/prince2.pdf
[accessed 11 January 2006].
Todd, GP; Markowitz, HM; (2000) Mean-Variance Analysis in
Portfolio Choice and Capital Markets; Frank J. Fabozzi
Associates.
Tseng, C; Gymytrasiewicz, P.J (2000) “Real Time Decision
Support System for Portfolio Management”, 35th Annual
Hawaii International Conference on System Sciences
(HICSS’02), Volume 3, January 07 – 10.
Turban, E; Aronson, AE (2001) Decision Support Systems and
intelligent systems, Sixth Edition, Prentice-Hall Incorporated.
Wooldridge. M (2002) An Introduction to Multiagent Systems,
John Wiley and Sons Ltd.
Country
Expected Return
USA
0.06634338
Global Equity Asset
Canada
0.070273699
Germany
0.02871881
UK
0.040776844
Japan
-0.043841495
Canada
Germany UK
Japan
Australia Thailand Colombia
Australia
0.053515971
0.03547
0.04672
0.02897
0.04817
0.04817
0.17519
0.07985
Thailand
-0.34793296
0.03593
0.04732
0.02934
0.04879
0.04879
0.17745
0.08088
Colombia
-0.026992001
0.04732
0.06232
0.03864
0.06426
0.06426
0.23370
0.10652
Brazil
-0.003288647
0.02934
0.03864
0.02396
0.03984
0.03984
0.14490
0.06604
0.04879
0.06426
0.03984
0.06626
0.06626
0.24097
0.10983
0.04879
0.06426Edward
0.03984
0.06626Dr Abel
0.24097
Falconer is0.06626
a research student;
Usoro, Dr0.10983
Mark
and Dr Brian
Lees are0.24097
his academic 0.87637
supervisory team.
0.17745
0.23370Stansfield
0.14490
0.24097
0.39945
0.08088
0.10652
0.06604
0.10983
0.10983
0.39945
0.18207
0.07624
0.10041
0.06226
0.10353
0.10353
0.37653
0.17162
12
Br
0.0
0.0
0.1
0.0
0.1
0.1
0.3
0.1
0.1
Semantic Databases: An Information Flow (IF) and Formal
Concept Analysis (FCA) Reinforced Information Bearing
Capability (IBC) Model
Yang Wang and Junkang Feng
Database Research Group
Semantic Database (SDB) seems hitherto somehow
overlooked in the literature compared with its ‘big
brother’, Semantic Web. What are the hindrances to
the development of SDB, which hence have to be taken
into account as we observe, include information
representation, knowledge management, meaning
elicitation, constraints/regularity identification and
formulation, and also partiality preservation. We
propose an architecture, which is a result of
reinforcing the notion of the Information Bearing
Capability (IBC) that we put forward elsewhere before
by applying the theory of Information Flow (IF) and
that of Formal Concept Analysis (FCA). We believe
that this architecture should enable SDB to cover a
number of these aspects, which build upon and go
beyond the relational database (RDB).
1. INTRODUCTION
Semantic Web (SW) is the supreme elegance of
topics, which covers numerous fields, such as
knowledge organization and management,
network technology and even data modeling.
Comparing to this prosperous triumph, the
seemingly evident lack of attention to Semantic
Database (SDB) would appear rather peculiar.
Whereas it is well known that SDB aims at
capturing, modeling and yielding meanings rather
than raw data, we observe that the short of robust
theoretical modeling foundation and guidance lies
as a gulf before the ‘fortune’. In our opinion, if
we want to achieve a satisfactory SDB, not only
primary pre-requisites such as capturing more
semantics and constraints, but also profound
concepts of information, representations and
partiality, need to be addressed.
To get across this gulf, the foundation of this
research is a series of theories (we refer to them
as ‘SIT’, short for Semantic Information
Theories) concerning semantic information and
information flow including Dreske (1981), Devlin
(1991), and in particular, Barwise and Seligman’s
(1991) information channel theory (IF for short).
We believe that an Information Flow (hereafter IF
for short) and Formal Concept Analysis (FCA)
reinforced Information Bearing Capability (IBC)
model (We will say more about it shortly)
provides a new prospective to SDB, which both
assures traditional requirements of design and
brings up some philosophical and mathematical
insights. This would, therefore, promote SDB to
be compatible with Knowledge base (KB) and
hence to be a strong support for SW.
1.1 A Short Review of Semantic Databases
(SDB)
A database system is a representation system,
which should be able to reflect real objects in the
circumstance being modeled. The content of a
database rests with what actually exists in the
modeled domain while any change operates on
this content should correspond with what happens
to those real world objects. Sustaining this tie is
not easy as at the first glance. Designing a data
model that captures as much as meaning as the
modeled domain is the solution of many
researchers (Hammer and McLeod 1981,
Jagannathan et.al, 1988, Tsur and Zamolo 1984).
To this end, concepts around SDB came into the
scene.
Bearing the goal of representing, describing and
structuring more semantics and meanings than
contemporary database (viz. Relational Database)
in mind, SDB needs to be closely related to the
modeled domain. Hammer addresses a number of
criteria that should be enforced during SDM
design (Hammer and McLeod 1981):

The constructs of the database model
should
provide
for
the
explicit
specification of a large portion of the
meaning of a database. So called semantic
expressiveness is not sufficiently achieved
by many current data modeling
techniques, such as hierarchical, network,
and relational models
13


A database model must support a relativist
view of the meaning of a database, and
allow the structure of a database to
support alternative ways of looking at the
same information. Being capable of
capturing more meaning requires never
rigid definitions and distinctions between
‘entities’, ‘attributes’ and ‘association’.
A database model must support the
definition of schemata that are based on
abstract entities. This point, in fact,
addresses that a database should have the
mechanism to support possible semantic
constraints.
In the related literature, there are mainly two most
interesting streams identified by the authors in
SDB modeling. The first one is that some of the
researchers are developing their SDM structure
on the root of available modeling techniques.
Most related to this research, some systems are
inheriting the basic modeling constructs of
RDM’s apparatus, for example, Iris Data Model
(Lyngback and Vianu 1987), Generic SDM (Chen
and McLeod 1989) and SDB management System
SIM (Boyed 2003). Meanwhile, Rishe and his
group build up a Semantic Wrapper over RDB
which produces set of SDB tools including
Knowledge database tool, Knowledge base and
Query Translator
(http://n1.cs.fiu.edu/SemanticWrapper.ppt). The
second is that some research shows that SDB is
more likely linked to Ontology and Knowledge
base
(http://www.fmridc.org/f/fmridc/dmt/sdm.html).
This would seem to orientate SDB to flourishing
the development of SW.
Besides this, currently, research around SDB
encounters numerous obstacles. The bottleneck,
as we have identified, resides in lack of certain
infrastructure to retrieve semantics and formulate
semantic constraints, not from traditional
database point of view but follow vigorous
guidance of Semantic Information Theory (SIT in
short). We believe that by philosophically
separating truly information from raw data, dually
grasping semantic constraints and partially
representing semantic information relation, an
advance model of SDB can be achieved.
1.2 IF and FCA Based IBC Prospect of SDB
In 1998, we identified a research problem, namely
the ‘information content’ of a formalized
information system (Feng 1998). In that paper
numerous works were cited and it was shown that
the main cause of this problem seemed that
information had been treated as ‘mystical liquid’.
We then argued that the lack of clearly expressed
and defined ‘information content’ of a conceptual
data schema was responsible for many difficulties
in data modeling and analysis as a process of
inquiry, which is a basis for the design of an
information system.
Then in 1999 we formulated a notion called
‘information bearing capability’ (IBC for short)
by drawing on interdisciplinary views of
information creation and transmission (Feng
1999). A four-facet principle currently elaborates
this notion, which is concerned with a set of
sufficient and necessary conditions for the IBC of
an information system. The conditions are:
information content containment,
distinguishability, accessibility and derivability
(Feng 2005). The principle about IBC and their
associated concepts that have been put forward in
a series of research papers (such as Xu and Feng
2002, Feng and Hu 2002, Xu 2005, and Wang
and Feng 2005a) may be seen as forming an
innovative perspective for looking at information
systems. Now, IBC as a cornerstone is applied to
a number of research problems that are being
looked at by our group such as schema mapping,
data exchanging and modeling. The ideas around
IBC however should be further developed and
tested in real world applications. To this end, it
seems that the most appropriate tool to reason
about and verify IBC would be IF combined with
FCA. We envisage that endeavor along this line
will uplift the articulation of what might be called
‘the microscopic infrastructure’ of the IBC
principle to an adaptable, adoptable and
applicable level in SDB modeling.
This paper proceeds as follows. In the next
section, we highlight some aspects of SDB
modeling that seem to have been overlooked in
the light of SIT rooted IBC model. Our approach
of combined use of IF and FCA in the IBC model,
which would, we believe, advance the state of the
art of SDB, is introduced in section 3. Following
this, a conceptual picture of IF and FCA
reinforced IBC model for SDB described and
elucidated in section 4.
2.
What Should a SDB
Represent and Provide?
Model,
14
As aforementioned, SDB is proposed in the
literature to address those problems encountered
in other forms of data modeling. As summarized
by Boyed (2003), there are several essential goals,
which need to be sustained, during SDB
development. The SDB is a high-level semanticsbased database description and structural
formalism for databases (Hammer, 1981).
Although attempting to capture all the semantics
of the modeled domain is unattainable, SDB
should endeavor to incorporate most of the
semantics. SDB advances RDB and other
database models in terms of its real-world
perception of the problems, different perspectives
of queries, and most importantly its inheritancebased hierarchical modeling structure. In addition
to these known characteristics, following the
insight of IBC based on SIT, we would propose
more significant features for SDB. Only when
these features are delivered can we say that SDB
is satisfiably achieved.
2.1 Data, Information and Semantics
Database is the vehicle for storing and providing
information. Without the guidance of
interdisciplinary philosophical semantic
information theory, it is not surprising that
contemporary database modeling dose not
separate data and truly information.
Notwithstanding modeling methods like RDB
being many and varied, as far as SDB is
concerned, it should broaden its edge to tackle the
truth of data, information, meaning and semantics
in order to capture semantics and to solve some
difficult issues, for example, query answering,
lossless transformation, etc.
In a typical contemporary database, ‘what you see
is what you get’ is the prevailing feature. Relation
between data and information remains
scrupulously bypassed. For a long decade, data
with its meaning is treated as information in the
context of database (Checkland, 1981). A famous
schema transformation approach, i.e.,
‘information capacity’ (IC) (Miller 1993),
straightly takes data instances of schemata as
information. Fusing Organizational Semiotics
(OS) into database, ‘meaning is created from the
information carried by signs’ (Mingers 1995).
A veritably practical SDB should take the
challenges that lie in several aspects around
definitions of information, information content
and meaning. Some of my colleagues have
provided an analysis about this (Wang and Feng
2005). Firstly, instances are not always faithful to
their semantic types. Traditionally, the schema of
a database is thought to represent the type level of
information while database instances fill into
these type level classes whereby receive their
semantics or meaning from the classes. However,
this view overlooks the facts that instances may
not loyal to their respective semantic
infrastructures. These instances do not represent
any information that originated the types (Dretske
1991). Secondly, the meaning of data in the
database is not necessarily to be part of their
information content. SDB should be able to use
alternative ways to represent the same
information. Therefore, a data construct
represents a piece of information only when the
information content of the data construct includes
that piece of information. It is not convincing to
use meaning as the criteria for the information
content of a piece of data. Finally, it is not
adequate to take the ability of accommodating
instances into the schema as the information
capacity of data constructs in the database (Wang
and Feng 2005). The fewer constraints being
modeled, the less specific the instances are.
Hence, less information there is. SDB modeling
should take this point into consideration and
facilitate it.
2.2 Constraints and Representations
No matter what form it is in; a database is after all
need to represent objects and relations in the
represented domain. The modes of representation
(Shimojima 1996) obey structural constraints that
mirror the regularities that govern things going on
in the represented domain. Any representation
involves certain kind of information flow.
Information flow results from the regularities in a
distributed system (Barwise and Seligman 1997,
P.8).
Contemporary database like RDB limit
themselves into a particular structure of
constraints such as relational objects and
associated relations. SDB should go beyond these
limits in the way of finding the best fit between
the representing system and the represented
domain. Apart from this aspect, SDB should also
ensure that its reasoning be consistent with the
represented domain. In other words, reasoning
over constraints needs great care. Wobcke (2000)
identifies the differences between schema-based
15
and information flow based reasoning. The
former is partly subjective and defeatable
contrasting to the objectiveness and nondefeatability holding by the latter. If given a fixed
context by discarding all alternative situations,
schema-based reasoning and information flow
based reasoning are transferable. Shimojima
(1996) uses basic mathematical instruments to
model constraints in order to perform a rigorous
investigation on a wide range representation
issues. His research provides a sound theoretical
foundation for developing our IBC model for
SDB in virtue of inferential reasoning intimate to
what happens in the domain to be modeled.
2.3 Partiality
Talking about semantics, it is evident to many
researchers, especially those who are familiar
with logics and linguistics, that there are ‘holes in
reality’ (Duzi 2003). These holes reside in our
abstract way of modeling particular dependency
relations among real world objects. Many
attempts have been made to philosophically
address such issues as Possible World Semantics
and Situation Semantics. As aforementioned, in
database, there exit instances that do not inherit
semantics from its corresponding class types.
Following Duzi (2003), if we take these instances
as the logical construction C (not unlike the
notion of ‘concept’ of Dretske 1981) for the
‘mode of representation’, which is discussed in
previous section, it should link the expression E
and its denotation D.
Problems arise when we use empty concepts, the
construction C will fail to achieve anything, not
even any meaning. As a result, the denotation D
will fail to give any truth-value in an argument.
Macroscopically, it is necessary for SDB to be
equipped with partial order to handle overall
informational relationships. Based on Dretske’s
information flow (1991) and Barwise and Perry’s
situation semantics (1993), Wobcke (2000)
argues that using conditionals as basic appliance,
people could evaluate the subjectiveness and
intentionality of a collection of schemata. The
idea is to treat those conditionals as expressing
constraints which are actually informational
relations between facts and events of the kind that
can be modeled using structures of situations
(Wobcke 2000). The order of situations for the
collection of constraints is in the form of partial
order supporting subjective reasoning. In certain
circumstances, i.e., providing certain fixed
context (situation), reasoning on this order is
identical to the reasoning of information flow.
Also, in Duzi’s thesis (2001), she points out that
information content inclusion relations (in
relation to attributes) are of partial order. Most
specifically, she formalizes informational
capability in a complete lattice based on the
power set of the attributes in question.
Furthermore, it is interesting that this lattice is
proved to be isomorphic to its substituting partial
ordered set of equivalence classes.
Therefore, for the sake of manipulating
informational scenarios, the need of supporting
partial order of the IBC model both
philosophically and mathematically should not be
ignored. Moreover, we believe that such a work
would be aligned with issues in knowledge
representation in the AI field.
3. Architecture based upon Information
Flow (IF) and Formal Concept
Analysis (FCA)
The central idea of IBC is called the IBC
principle. This principle, is made up of conditions
of information content containment,
distinguishability, accessibility and derivability
and it is put forward by Feng (2005) and his
colleagues through a period of arduous work in
the sense of drawing interdisciplinary views of
information creation and transmission (Feng
1999, Xu and Feng 2002, Feng and Hu 2002, Xu
2005, and Wang and Feng 2005a). IF is first
introduced into IBC for reasoning about and for
verifying the principle (Wang and Feng 2005). As
being successively compatible and content with
the IBC, IF has become a headstone for further
development and application of the IBC model.
For the purpose of elevating implementation,
FCA is probed and found that it is adaptable,
applicable and adoptable both theoretically and
practically with IF.
3.1 Channel-Theoretical Information Flow
The Channel-theoretical Information Flow theory
(IF) is a mathematical model of semantic
information flow. Information flow is possible
due to the regularities among normally disparate
components of a distributed system. It is known
that such a theory succeeds in capturing partial
16
order of classifications (Kalfoglou and
Schorlemmer 2005) that underlies the flow of
information. Sophisticated notions (we do not go
into details here) stemming from IF now have
been formulated for explorations on semantic
information and knowledge mapping and
exchanging. Kent (2002a, 2002b) exploits
semantic integration of ontologies by extending a
first order logic based approach (Kent 2000)
which is also based on IF. An information flow
framework (IFF) has been advocated as a metalevel framework for organising the information
that appears in digital libraries, distributed
databases and ontologies (Kent 2001). From
Kent’s work, Kalfoglou and Schorlemmer
(2003a) develop an automated ontology mapping
method in the field of knowledge sharing and
cooperation. IF and its surrounding concepts are
also relevant to solving problems of semantic
interoperability (Kalfoglou and Schorlemmer
2003b). Apart from this main stream of
applications, IF supports various research efforts
from defensible reasoning (Cavedon 1998); endoperspective formal model (Gunji et al 2004) to
semiconcept and protoconcept graphs (Malik
2004).
Besides the effective effort of using IF to
represent, capture and model constraints for a
given modelled domain, it is also observed that IF
‘was not developed as a tool to be used in real
world reasoning’ (Devlin 1999) and we observe
that it is on its own insufficient for describing
domain information or knowledge. To fill these
gaps, Formal Concept Analysis (FCA) was
proposed as a silver bullet.
3.2 Formal Concept Analysis (FCA)
FCA was developed by Rudolf Wille (Wille
1982) as a method for data analysis, information
management, and knowledge representation (Priss
2005a). Presumably due to its applicable nature, it
does not take long for FCA to become a common
interest in many research communities, for
example, social net work analysis (Freean and
White 1993), linguistics (Priss 2005b), and
software engineering (Fischer 1998, Eisenbarth et
al. 2001). As aforementioned, FCA provides solid
foundations for not only information and
knowledge retrieval by its underlying
mathematical theory (Godin et al. 1989,
Kalfoglou et al. 2004) but also for respective
representations by concept lattice (Wille 1982,
1992, 1997b) along with concept graphs (Prediger
and Wille 1999). We maintain that the use of
FCA will supplement with IF in SDB modeling.
By using IF along, it would appear that the
construction of an ‘information channel’ in many
cases is difficult when applying IF to real
information system problems. To alleviate it, we
envisage that ‘Conceptual Scaling’ techniques
(Ganter and Wille 1989, Prediger and Stumme
1999)’, which are affinity with FCA, will be
useful. Furthermore, reasoning and inference over
difference levels of a channel can be
characterized by ‘Concept Graph’ (Prediger and
Wille 1999) in the light of FCA-based ‘Concept
Lattice’ (Wille 1982, Wille 1992, Wille 1997b).
In other words, FCA provides the investigation
with a basis for extraction, representation and
demonstration of informational aspect of
semantics, and at the same time IF-based
techniques/methods can be charged with the task
of information flow based reasoning. As a result,
the combined use of IF and FCA can shed some
light on solving problems around the IBC within
the context of SDB, which is also harmonious
with knowledge discovery and representation.
3.3 Prospect of Combined Use of IF and FCA
The essential element of our IBC mode for SDB
is the combined use of IF and FCA. They provide
vital insights for our SDB model. The
compatibility between them is crucial for any
combined use. We give reasons below for using
IF theory and the theory of FCA in combination.
Firstly, both IF and FCA share the same origin,
i.e., category theory with the means of Chu space
(Gupta 1994, Barr 1996 and Pratt 1995). As
Wolff (2000) observes, ‘it is really astonishing
that these tools (IF and FCA) are not mutually
taken into account in each other’s theory’. Priss
(2005a) treats the ‘classifications’ in IF as a
general sense of ‘concept lattices’ in FCA.
Following this line of thinking, secondly, nearly
all fundamental concepts invented by both of IF
and FCA can find counterparts in each other. For
example, the notions of ‘classifications’ in IF
matches that of ‘formal context’ in FCA;
‘information channels’ in IF matches ‘scaled
many-valued contexts’ in Conceptual Scaling
(Ganter and Wille 1989, Ganter and Wille 1999)
associated with FCA. Other basic notions
presented in IF, such as ‘state space’, ‘refinement
of channels’, and ways of handling ‘vagueness’
17
are also delivered in FCA mathematically (Wolff
2000). Finally, IF bears epistemological
resemblance to FCA. To be explicit, starting from
the same algebraic category, IF together with
FCA aim at formulating and justifying ‘partial
order’ that relies on agreed understanding of the
existence of ‘duality’ between separated
situations, which is exactly why information flow
commences.
Combined use of IF and FCA is beneficial to
constructing the IBC model of SDB. SDB highly
needs to capture more semantics. In IF and FCA
reinforced IBC model, FCA would serve as the
linkage between IF reasoning and the modelled
domain. Due to the ‘non-directly-applicable’
nature of IF (Devlin 1999), applying it directly to
modeling informational semantics proves to be
problematic. In contrast, a number of works
stemming from FCA around knowledge discovery
and information retrieval have been put forward.
For example, Stumme and his colleagues have
encouraged the use of FCA in exploration and
representation of implied information and
facilitating the conversion of information into
knowledge (Hereth et al. 2000, Stumme et al.
1998). We would use the ‘Conceptual Scaling’
techniques (Prediger and Stumme 1999, Prediger
and Wille 1999) to combine FCA with IF
reasoning because of FCA’s logical equivalence
with ‘Information Channel’. The results of
reasoning would be presented in Concept Graphs,
which has advantages in representing semantics
in partial order.
Also, a combined use of IF and FCA can
satisfactorily model more semantic constraints
identified by Hammer (1987). To tackle
information flow, IF insists on analyzing relations
between tokens and types. According to the
second principle of information flow, i.e.,
‘information flow crucially involves both types
and their particulars’ (Barwise and Seligman
1997, P.27). Originally and largely following
Dretske (1981), we thought that semantics are
presented on the type level which further provides
the meanings to the tokens involved in
information flow. However, from the paper of
Kalfoglou and Schorlemmer on IF-map (2003a),
we find the important role of tokens, e.g., the
same set of rivers and streams, played in
determining semantics or constraints of the whole
system in terms of semantic correspondences
between the types. We observe that in fact,
Kalfoglou and Schorlemmer has employed
primary thinking of FCA in exploring ‘intension’
and ‘extension’ of formal concepts within a given
formal context. That is from either set, i.e.,
intensions or extensions; we can define its
counterpart in the context, and thus the formal
concepts. Therefore, using relations in tokens
(extensions), we would gain relation of concepts
and hence arrive at a set of constrains, which
reflect a type of regularities of the whole system
in the given context. This is exactly how tokens
take part in defining the semantics of a system,
and in achieving semantic interoperability.
Further to this point, we envisage that duality held
by both IF and FCA enables us to support
alternative ways for the user to view even the
same information in SDB. Start with the relations
that reside in types and we would end up with
relation of tokens and vice versa. Therefore,
depending on what aim we want to achieve, we
could selectively take either tokens or types as
our starting point in different analysis. Explicitly,
if we want to solve the semantic interoperability
problem, as Kalfoglou and Schorlemmer did, we
shall investigate tokens-determined relations in
order to achieve the relations on types. On the
other hand, if we want to find out why and how
data constructs represents (or conveys) the
information about a given semantic relation (i.e.,
a relation between some real world objects), in
most cases, we will take the semantics on types of
this structure as a foundation.
4. Outline of IF and FCA Reinforced
IBC Model for SDB
Based on previous sections, we can now start
describing the IF and FCA reinforced IBC model
designed for SDB. We will begin with data
schemata as we believe that original databases
and schemata is too valuable to be retained
(Figure 1).
The original database schema together with a
serial of dependencies held by the schema would
be analyzed by using IF and FCA. This analysis
needs to be assisted by obtained initiative
business constraints e.g. stake holder views,
presented in the format of scales, so that
subjectiveness is preserved at this early stage. The
construction of ‘information channel’ of IF will
benefit from the technique of ‘conceptual scaling’
of FCA. The output of investigation is a
18
conceptual space which contains all the
constraints (semantics) captured by every
information channel. This space is called by us as
the ‘kernel of IBC’. When the user puts a query
for a piece of information to this kernel, if there is
no direct answer, an inference will be carried out
by means of a set of ‘information content
inference rules’ (Feng and Hu 2002). Then, final
results are added into a separate conceptual space
following the decision of the user. Connected
with knowledge representation and management,
the consequent results could be transformed using
XML-extended Information Flow Framework
(IFF)
(http://www.ontologos.org/IFF/The%20IFF%2
0Language.html) language.
Figure 1. Overall Picture of IF and FCA Reinforced IBC model
There are two most important parts of this model
to construct corresponding IF channels. The
which show in two boxes in Figure 1. To clarify
many-valued context will then become singlewhat actually happens inside of them, we will use
valued context as a result. Following this, using
two more diagrams.
dependencies that are determined by business
In Figure 2, there is a detailed process for arriving
rules as the other scales, another scaling, i.e., the
at the kernel of IBC. Both primary database
‘relational scaling’, will be accomplished by a
schemata and instances are translated into manyfinal lattice layout also with a crowd of
valued context by FCA. Then, two scaling
information channels. The ultimate results are sets
processes are performed. The first one called
of ‘IF’ theories derived from all of the channels.
‘conceptual scaling’. It is based on the idea that
This is what we want to model as the system
embedded structural constraints are used as scales
regularities.
19
Figure 2. How to Achieve Kernel of IBC
In addition, another significant part in our model
(Wang 2005 and Xu 2005), we found that these
is inference on information content (Figure 3).
inference rules can be justified by theory of IF.
The information content based inference rules
In the future, we will generalize these
are put forward by Feng and Hu (Feng and Hu
verifications by not only IF but also FCA.
2002). Furthermore, through two MSc projects
20
Figure 3. Information Inference Rules (IIR)
21
Computing and Information Systems
© University of Paisley 2006
5. CONLUSIONS
This paper represents our first step towards
satisfactorily modeling a SDB by means of an IF
and FCA reinforced IBC. Three more criteria,
i.e., extracting information, modeling semantic
constraints and also partially representing
information flow, have been proposed in
addition to traditional SDB requirements. The
overall idea of the IBC model for SDB is shown
with diagrams that heavily draw on concepts
from both IF and FCA. This attempt seems
worthwhile for the development of SDB, and it
is also compatible with most modern knowledge
management systems, and therefore relevant to
the area of semantic web.
References
Barr, M. (1996). The Chu construction. Theory and
Applications of Categories, 2(2):17–35.
Barwise, J. and Seligman, J. (1997) Information
Flow: the Logic of Distributed Systems,
Cambridge University Press, Cambridge.
Barwise, J. and Perry J. (1983). Situations and
Attitudes, Cambridge, Mass.: Bradford-MIT.
Boyed, S. (2003). A Semantic Database Management
System: SIM. The University of Texas at Austin,
Department of Computer Sciences. Technical
Report CS-TR-03-43.
Cavedon, L. (1998). Default Reasoning as Situated
Monotonic Inference, Minds and Machines 8.
Checkland, P. (1981). Systems Thinking, Systems
Practice. Chichester, UK: Wiley.
Chen, I. A. and McLeod, D. (1989). Derived Data
Update in Semantic Databases, Proceedings of
the fifteenth international conference on Very
large data bases, p.225-235, July, Amsterdam,
The Netherlands.
Devlin, K. (1991). Logic and Information,
Cambridge.
Devlin, K. (2001). Introduction to Channel Theory,
ESSLLI 2001, Helsinki, Finland.
Dretske, F. (1981). Knowledge and the Flow of
Information, Basil Blackwell, Oxford.
Duží, M. (2001). Logical Foundations of Conceptual
Modelling. In VŠB-TU Ostrava.
Duží, M. (2003). Do we have to deal with partiality?.
In Miscellania Logica, vol. Tom V, 45-76.
Eisenbarth, T., Koschke, R., & Simon, D. (2001).
Feature-driven Program Understanding using
Concept Analysis of Execution Trace. In
Proceedings of the Ninth International Workshop
on Program Comprehension. International
Conference on Software Maintenance.
Feng, J. (1998). The "Information Content’ Problem of
a Conceptual Data Schema, SYSTEMIST, Vol.20,
No.4, pages 221-233, November 1998. ISSN:
0961-8309
Feng, J. (1999). An Information and Meaning Oriented
Approach to the Construction of a Conceptual
Data Schema, PhD Thesis, University of Paisley,
UK.
Feng, J. and Hu, W. (2002). Some considerations for a
semantic analysis of conceptual data schemata, in
Systems Theory and Practice in the Knowledge
Age. (G. Ragsdell, D. West. J. Wilby, eds.),
Kluwer Academic/Plenum Publishers, New York.
Feng, J. (2005). Conditions for Information Bearing
Capability, Computing and Information Systems
Technical Reports No 28, University of Paisley,
ISSN 1461-6122.
Fischer, B. (1998). Specification-Based Browsing of
Software Component Libraries. Proc. Automated
Software Engineering, Hawaii, 246-254.
fMRIDC,
The
Semantic
Database
Model,
http://www.fmridc.org/f/fmridc/dmt/sdm.html.
Ganter, B. and Wille, R. (1999). Formal Concept
Analysis: mathematical foundations. Springer.
ISBN: 3-540-62771-5.
Godin, R., Gecsei, J., & Pichet, C. (1989). Design of
Browsing Interface for Information Retrieval. In N.
J. Belkin, & C. J. van Rijsbergen (Eds.), Proc.
SIGIR ’89, 32-39.
Gupta, V. (1994). Chu Spaces: A Model of
Concurrency. PhD thesis, Stanford University,
1994.
Hammer, M., McLeod, D. (1981). Database
Description with SDM: A Semantic Database
Model, ACM Trans. Database Syst. 6 (3): 351-386.
Hereth, J., Stumme, G., Wille, R. and Wille, U.
(2000). Conceptual Knowledge Discovery in Data
Analysis. In B. Ganter, & G. Mineau (Eds.),
Conceptual Structures: Logical, Linguistic and
Computational Issues. LNAI 1867. Berlin:
Springer, 421-437.
Information Flow Framework (IFF) Language,
http://www.ontologos.org/IFF/The%20IFF%20Lan
guage.html.
Jagannathan, D., Guck, R. L., Fritchman, B. L.,
Thompson, J. P., Tolbert, D. M. (1988). SIM A
Database System Based on the Semantic Data
Model. SIGMOD Conference: 46-55.
Kalfoglou, Y. and Schorlemmer, M. (2003a). IFMap:
an
Ontology
Mapping
Method
based
onIinformation Flow Theory. Journal on Data
Semantics, 1(1):98–127.
Kalfoglou, Y. and Schorlemmer, M. (2003b) Using
Information Flow Theory to Enable Semantic
Interoperability, In Proceedings of the 6th Catalan
Conference on Artificial Intelligence (CCIA '03),
Palma de Mallorca, Spain, October 2003.
Kalfoglou, Y., Dasmahapatra, S., & Chen-Burger, Y.
(2004). FCA in Knowledge Technologies:
Experiences and Opportunities. In P. Eklund
(Ed.), Concept Lattices: Second International
Conference on Formal Concept Analysis, LNCS
2961. Berlin: Springer, 252-260.
Kalfoglou, Y., Schorlemmer, M. (2005). Using
Formal Concept Analysis and Information Flow
for Modeling and Sharing Common Semantics:
lessons learnt and emergent issues, In Proceedings
of the 13th International Conference on
Conceptual Structures (ICCS2005), Kassel,
Germany, July 2005
Kent, R. E. (2000). The Information Flow
Foundation
for
Conceptual
Knowledge
Organization. In: Dynamism and Stability in
Knowledge Organization. Proceedings of the
Sixth International ISKO Conference. Advances
in Knowledge Organization 7 111–117. Ergon
Verlag, Würzburg.
Kent, R. E. (2001). The Information Flow
Framework. Starter document for IEEE P1600.1,
the IEEE Standard Upper Ontology working
Group, http://suo.ieee.org/IFF/.
Kent, R. E. (2002a.) The IFF Approach to Semantic
Integration. Presentation at the Boeing MiniWorkshop on Semantic Integration, 7 November
2002.
Kent, R. E. (2002b). Distributed Conceptual
Structures. In: Proceedings of the 6th
International Workshop on Relational Methods in
Computer Science (RelMiCS 6). Lecture Notes in
Computer Science 2561. Springer, Berlin.
Kollewe, W., Skorsky, M., Vogt, F., and Wille, R.
(1994). TOSCANA – ein Werkzeug zur
begrifflichen Analyse und Erkundung von Daten.
In R. Wille, &19 M. Zickwolff (Eds.),
Begriffliche Wissensverarbeitung - Grundfragen
und
Aufgaben.Mannheim:
B.I.Wissenschaftsverlag, 267-288.
Lyngbaek, P. and Vianu, V. (1987). Mapping a
Semantic Database Model to the Relational
Model, Proceedings of the 1987 ACM SIGMOD
international conference on Management of data,
p.132-142, May 27-29, San Francisco, California,
United States.
Malik, G. (2004.) An Extension of the Theory of
Information
Flow
to Semiconcept
and
Protoconcept Graphs. ICCS 2004: 213-226.
Miller, R. J., Ioannidis, Y. E. and Ramakrishnan, R.
(1993). The Use of Information Capacity in
Schema Integration and Translation, in
Proceedings of the 19th International Conference
on Very Large Data Base, Morgan Kaufmann, San
Francisco.
Mingers, J (1995). Information and Meaning:
Foundations for an Intersubjective Account.
Journal of Information Systems 5 285-306.
Pratt, V. (1995). The Stone gamut: A coordinatization
of mathematics. Logic in Computer Science, pages
444–454.
Prediger, S. and Stumme, G. (1999). Theory-driven
Logical Scaling. Conceptual Information Systems
meet Description Logics. In P. Lambrix, A.
Borgida, M. Lenzerini, R. Muller, & P. PatelSchneider (Eds.), Proceedings DL’99. CEUR
Workshop Proc. 22.
Prediger, S. and Wille, R. (1999). The Lattice of
Concept Graphs of a Relationally Scaled Context.
In W. Tepfenhart, & W. Cyre (Eds.), Conceptual
Structures: Standards and Practices. Proceedings of
the 7th International Conference.
Priss, U. (2005a). Formal Concept Analysis in
Information Science. Annual Review of
Information Science and Technology. Vol 40.
Priss, U. (2005b). Linguistic Applications of Formal
Concept Analysis. In: Ganter; Stumme; Wille
(eds.), Formal Concept Analysis, Foundations and
Applications. Springer Verlag. LNAI 3626, p. 149160.
Rishe, N., Semantic Wrapper over Relational
Database.
http://n1.cs.fiu.edu/SemanticWrapper.ppt.
Shimojima, A. (1996). On the Efficacy of
Representation, Ph.D. Thesis. The Department of
Philosophy, Indiana University.
Stumme, G., Wille, R., and Wille, U. (1998).
Conceptual Knowledge Discovery in Databases
using Formal Concept Analysis Methods. In J. M.
Zytkow, & M. Quafofou (Eds.), Principles of Data
Mining and Knowledge Discovery. LNAI1510.
Berlin: Springer, 450-458.
Tsur, S. and Zaniolo, C. (1984) An implementation of
GEM Supporting a Semantic Data Model on a
Relational Back-end, Proceedings of the 1984
ACM SIGMOD international conference on
Management of data, June 18-21, Boston,
Massachusetts.
Wang, X. and Feng, J. (2005b.) The Separation of
Data and Information in Database System under an
Organizational Semiotics Framework. The 8th
International
Workshop
on
Organizational
Semiotics, Toulouse, France.
Wang, Y. and Feng, J. (2005a). Verifying Information
Content Containment of Conceptual Data
Schemata by Using Channel Theory. The 14th
International Conference on Information Systems
Development, Karlstad, Sweden. Springer-Verlag.
23
Wille, R. (1982). Restructuring lattice theory: an
Approach based on Hierarchies of Concepts. In I.
Rival (Ed.), Ordered sets. Reidel, DordrechtBoston, 445-470.
Wille, R. (1992). Concept Lattices and Conceptual
Knowledge Systems. Computers & Mathematics
with Applications, 23, 493-515.
Wille, R. (1997a). Conceptual Graphs and Formal
Concept Analysis. In D.Lukose, H. Delugach, M.
Keeler, L. Searle, & J. F. Sowa (Eds.), Conceptual
Structures: Fulfilling Peirce’s Dream. Proc.
ICCS’97. LNAI 1257. Berlin:Springer, 290-303.
Wille, R. (1997b). Introduction to Formal Concept
Analysis. In G. Negrini (Ed.), Modelli e
modellizzazione.
Models
and
modelling.
Consiglio Nazionale delle Ricerche, Instituto di
Studi sulli Ricerca e Documentazione Scientifica,
Roma, 39-51.
Wobcke, W. (2000). An Information-Based Theory of
Conditionals. Notre Dame Journal of Formal
Logic 41(2): 95-141.
Wolff, K. E. (2000). Information Channels and
Conceptual Scaling. In Working with Conceptual
Structures. Contributions to ICCS 2000, Shaker
Verlag.
Xu, H. and Feng, J. (2002). ‘The "How" Aspect of
Information Bearing Capability of a Conceptual
Schema at the Path Level’. The 7th Annual
Conference of the UK Academy for Information
Systems, UKAIS'2002. Leeds . ISBN 1-898883149, pp.209-215
Xu, Z. (2005). Verifying Information Inference Rules
by using Channel Theory, MSc dissertation,
University of Paisley.
Wang.Y is a Researcher and Dr. Feng J. a Senior
Lecturer at the University of Paisley
24
Computing and Information Systems
© University of Paisley 2006
An Investigation into Trust as an Antecedent to Knowledge
Sharing in Virtual Communities of Practice
Abel Usoro, Mark W Sharratt and Eric Tsui
Information Systems Group
This study focuses on the role of trust in knowledge
sharing within the context of virtual communities of
practice. Trust is widely accepted as an important enabler
of knowledge management (KM) processes.
We
conceptualise trust across three dimensions, namely:
competence, integrity, and benevolence; we test
hypotheses as to the effect of these facets of trust on
knowledge sharing by surveying an intra-organisational
global virtual community of practitioners. The results
indicate that all three dimensions of trust are positively
related to knowledge sharing behaviour. Trust based on
the perceived integrity of the community was found to be
the strongest predictor of knowledge sharing behaviour.
Our findings suggest that the dimensions of trust buttress
each other; although they are theoretically distinct, they
appear to be empirically inseparable. We propose that in
order for knowledge sharing to be enabled, trust must
concurrently exist in all three dimensions. Implication to
organisations in their recruitment policy is to include
competence, integrity and benevolence in their sought-for
attributes of new employees. KM practitioners also have
to encourage these attributes in existing employees, who
are potential members of on-line communities of practice.
1
INTRODUCTION
The management of the knowledge base of
organisations is becoming an area of
strategic focus for many knowledgeintensive organisations (Ruggles 1998;
Beaumont and Hunter 2002). Indeed,
knowledge is unique as an organisational
resource in that most other resources tend
to diminish with use; the potential for
growth in knowledge resources increases
with use, as “ideas breed new ideas, and
shared knowledge stays with the giver
while it enriches the receiver” (Davenport
& Prusak 1998: p. 16-17).
Hence,
understanding
key
knowledge
management (KM) processes is a growing
area of organizational research efforts
(Scarbrough et al. 1999).
Within the KM field, there is
widespread acceptance of the role of
communities of practice as a key KM
enabler (Wenger et al. 2002; Scholl et al.
2004). CoPs have been described as
groups of people “informally bound
together by shared expertise and passion
for a joint enterprise” (Wenger and Snyder
2000: 139), “playing in a field defined by
the domain of skills and techniques over
which the members of the group interact”
(Lesser and Storck 2001: 831).
Such
communities provide a rich locus for the
creation and sharing of knowledge both
within and between organisations (Brown
and Duguid 1991; Lesser and Everest
2001). Information and communication
technologies are now extending the
boundaries of traditional face-to-face
communities
by
creating
virtual
communities
that
enable
global
asynchronous and real-time collaboration
(Hildreth and Kimble 2000).
These
communities may exist primarily, or solely,
online and become largely dependent on
computer-mediated communication to
connect members.
However, the availability of information
systems does not automatically induce a
willingness to share information and
develop new knowledge (Nahapiet and
Ghoshal 1998). Indeed, research has found
that social interaction that is mediated by
online technologies can be less effective
than in face to face meetings (Preece 2000).
Despite such limitations of technology,
research has shown that emotional
attachments can develop online, despite a
lack of face-to-face contact (Rheingold
1993; Sharp 1997; Preece 2000). The
building of trust is an important social
process that is widely accepted as a
prerequisite to cooperation (for example:
Gambetta 1988; Ring and Van de Ven
1994; Mayer et al. 1995; Nahapiet and
Ghoshal 1998; Wang and Ahmed 2003).
Research has shown that levels of trust
influence the exchange of resources
between intra-organisational business units
(Tsai and Ghoshal 1998) and research
investigating knowledge sharing has found
trust to be important in the receipt of
useful knowledge (Levin et al. 2002). It
logically follows that virtual communities
that fail to develop trusting relations will
restrict the development of knowledgesharing activities.
While Levin et al.’s research focused
on the receipt of knowledge, there appears
to be little current understanding as to the
importance of trust in the provision of
knowledge in general, and specifically
within the context of virtual communities
of practitioners. Ardichvili et al. (2003)
conducted an exploratory study into the
factors that affect knowledge sharing
within
intra-organisational
virtual
communities of practice. Their findings
identify a lack of understanding of the
role of trust in the provision of knowledge.
Our paper takes up this challenge by
investigating the role of trust, in its many
guises, in the provision of knowledge
within
a
virtual
community
of
practitioners.
The following section sets out our
operational definitions of knowledge
sharing and trust. This leads to the
development of a number of hypotheses as
to the relationship between three facets of
trust and knowledge sharing. We discuss
the design of the empirical research, and
test the hypothesised relationships.
Findings are then presented and reflected
upon. We conclude the paper with future
research directions and discuss the
limitations of this study.
2
OPERATIONAL DEFINITIONS
Knowledge is an intangible resource and is thus
difficult to measure. Indeed, a review of
the literature has established that
knowledge sharing in not well defined for
the purposes of empirical research. In
order to understand knowledge sharing, it
is necessary to define what we mean by
knowledge, and how this relates to
information and data.
2.1
Knowledge, Information and Data
Knowledge can only exist in the mind of the
individual (Van Beveren 2002). It is
through knowledge that we perceive the
world around us; knowledge largely
determines how we interpret and respond
to external stimuli. Hence, knowledge
often determines action. Knowledge is
acquired through a process of action and
reflection (Argyris & Schon 1978). Within
this
process,
the
interaction
and
communication of two or more individuals
can lead to the exchange and sharing of
knowledge. Here, information facilitates
this process and acts as a communicatory
representation of knowledge.
Data is the raw component of
information. Intrinsically, data contains no
meaning, data becomes information when
framed within a meaningful context. On
their own, the numbers 56 and 1648 are
just items of data. Framed within a
context,
such
data
may
provide
information, for example, the number 56
bus is due at 16:48 hours. Hence data is
transformed into information.
It is
knowledge which provides the context that
creates information from data and it is
through the interpretation of such
information that new knowledge may be
acquired.
Hence, information and
knowledge intrinsically facilitate the
knowledge-sharing process.
2.2
Knowledge Sharing
As we have described, knowledge per se cannot
exist outside the mind of the individual.
Knowledge sharing involves a process of
communication whereby two or more
parties are involved in the transfer of
knowledge. This is a process that involves
the provision of knowledge by a source,
followed by the interpretation of the
communication by one or more recipients.
The output of the process is the creation of
new knowledge. Hence, knowledge sharing
is defined as a process of communication
between two or more participants involving
the provision and acquisition of knowledge.
Indeed, the communication can take many
forms, using both verbal and non-verbal
mechanisms, with or without the use of
technology.
2.3
Trust
Trust is a concept about which there is much
debate (Kramer & Tyler, 1996). Amongst
the scientific community, the definition of
the concept lacks consensus (Fisman and
Khanna 1999; Adler 2001), although there
appears to be agreement that the construct
of trust is both complex and multifaceted
(Ford, 2003; Simons 2002). Fukuyama
(1995) views trust as “the expectation that
arises within a community of regular,
26
honest, and cooperative behaviour, based
on commonly shared norms, on the part
of the members of the community" (p. x).
This view of trust based on the
expectation of honest and cooperative
actions is shared by many (for example,
Gambetta
1988;
Mishra
1996;
Bhattacherjee 2002). Mayer et al. (1995)
define trust as “the willingness of a party
to be vulnerable to the actions of another
party based on the expectation that the
other will perform a particular action
important to the trustor, irrespective of
the ability to monitor or control that other
party” (p. 712). Both definitions above
relate to a confidence in others’ future
actions, with Mayer et al. extending their
definition by arguing that in order to be
vulnerable, one must be willing to take a
risk based on the trusting relationship.
An important distinction is that
Fukuyama views trust as based on shared
norms within a group; Mayer et al.
approach their definition from a dyadic
level, analysing the existence of trust
between two individuals.
Simons (2002) identifies Mayer et al.’s
(1995) definition of trust as often cited in
the literature. Mayer et al. identify three
Benevolence
attributes of another party in which
perceptions of trust can be based, namely,
benevolence, integrity and ability. This can
be seen to extend Fukuyama’s definition of
trust by incorporating ability as a specific
antecedent to trust. Mishra (1996) goes
further by defining four dimensions of
trust: concern, reliability, competence and
openness. While concern, reliability and
competence can be seen to mirror
benevolence,
integrity
and
ability,
respectively, Mayer et al. (1995) highlight
the overlap of the openness dimension with
their benevolence and integrity dimensions,
arguing that “Mishra's openness is
measured through questions about both the
trustee's general openness with others and
openness with the trustor, which could be
expected to be related to either integrity or
benevolence, respectively" (p. 722). This
overlap is supported by McKnight et al.
(2002) who presented confirmatory factor
analysis. Figure 1 presents the overlap
(mapping) of the two definitions of trust.
Given the discussion in this section, and in
the interests of parsimony, we adopt
Mayer’s conceptualisation of the three
characteristics of another party in which
trust may be held.
Concern
Reliability
Integrity
Competence
Ability
Mayer et al
Openness
Mishra
Figure 1: Mapping between Mayer et al. (1995) and Mishra’s (1996) definitional components of Trust
3 CONCEPTUAL MODEL: TRUST AS AN
ANTECEDENT TO KNOWLEDGE SHARING
In virtual communities, trusting relations can
emerge without any direct social
interaction due, in part, to the
transparency of online communications. A
newcomer typically has access to an
electronic record of previous discussions and
access to knowledge-based assets held in the
community’s
common
repositories.
Perceptions of trustworthiness based on
27
competence, honesty, benevolence and
behavioural reliability provide confidence
in future actions, and can be fostered by
the high degree of openness and visibility
surrounding
online
communications
within virtual communities. This in turn
fosters greater levels of cooperation, and
discourages
opportunistic
behaviour
(Fishman & Khanna, 1999).
Ability- or competence-based trust
exists when an individual believes that
another party has knowledge and
expertise in relation to a specific domain
(Jarvenpaa et al. 1998). This facet of trust
can be related to the fear of losing face
that Ardichvili et al. (2003) identified as
one of the main barriers to knowledge
sharing in online communities of practice.
For example, if a member’s perception of
her own competence is significantly lower
than the level of competence that she
associates with the virtual community, then
the motivation to publicly share her
knowledge may be affected due to the fear of
criticism or ridicule. This suggests a causal
link between one’s perceptions of the
community’s ability and engagement in
knowledge sharing, whereby high levels of
competence-based trust could restrict the
knowledge shared with a community.
However, whereas this argument appears to
make logical sense, the converse does not
logically hold. Where a member perceives a
community to be of low competence, such a
perception is unlikely to encourage
knowledge sharing.
In fact, such
perceptions are more likely to discourage
any form of voluntary participation in the
community.
28
Computing and Information Systems
© University of Paisley 2006
Another dynamic emerges when we
consider the perspective of competencebased trust in motivating community
participation. According to Lave and
Wenger’s (1991) theory of situated
learning, a newcomer to a community of
practice becomes involved in a transition,
over time, from peripheral participation in
the practice towards becoming a masterful
practitioner.
This process involves a
member learning by becoming situated
within the field of the community’s
practice. Within virtual CoPs, the process
of situating in the community’s practice
involves members creating and sharing
knowledge by engaging in intellectual
exchange through their participation in the
community’s
computer-mediated
communications.
By sharing and
developing ideas, by testing and validating
assumptions, by discussing, problem
solving and generally striving to become
more
competent
practitioners,
the
community members are able to engage in
the mutual development of both their own
knowledge and the community’s pool of
expertise.
With this ongoing process,
members engaged in the development of
cooperative and trusting relations whilst
simultaneously developing knowledge of
what it means to be a competent and
masterful practitioner (Nahapiet and
Ghoshal 1998). Within this process, we
define a member’s passion for the practice,
as the desire to become a more competent
practitioner and to engage in the
community’s practice. Where a member’s
passion for the practice is high, the
member will be more likely to seek
engagement and interaction with a
community of competent practitioners.
From this perspective, the role of
competence-based trust can be seen both as
a potential enabler and barrier to
community participation. When motivated
by a passion for the practice, sharing
knowledge with a community one perceives
as highly competent becomes an
intrinsically rewarding experience. The
rewards derived from the consensual
validation received from a community held
in high regard can act as a motivator to
members to share their ideas, thoughts,
and insights.
MacMillan and Chavis (1986) discuss
the process of consensual validation,
describing how “people will perform a
variety of psychological gymnastics to
obtain feedback and reassurance that ...
what they see is real and that it is seen in
the same way by others” (p. 11). In
communities of practice, consensual
validation may go further, acting as a
mechanism that represents a member’s
transition from peripherality to central
participation; in effect validating a
member’s standing within the community.
Thus, the community’s consensual
validation may act as a form of
recognition, establishing and confirming
one’s status as a knowledgeable
practitioner.
Conversely, sharing knowledge with a
community one perceives to be of low
competence will be an inherently less
rewarding experience. Not only is the
value derived from the community’s
validation reduced, member’s perception
of their self worth is also diminished by
identifying and participating within a
community perceived to be of limited
competence.
From the above discussion, we have
identified a different dynamic that the role
of competence-based trust may play in
terms of enabling and disabling
community participation in general, and
knowledge sharing specifically.
It is
possible that the motivating effect driven
by the opportunity to become a competent
practitioner could potentially overcome
the fear of losing face identified by
Ardichvili et al. (2003). Likewise, it is
equally possible that fear of losing face
will be the dominant force. In other
words, we can consider fear of losing face
as a moderating variable such that if it
exists at intense levels could change from
positive to negative the relationship
between competence and knowledge
sharing. Based on this uncertainty, we
propose the following hypothesis.
H1: One’s degree of trust in the competence
of a community is positively related to one’s
engagement in knowledge sharing with the
community.
A shared interest in the community’s practice can
foster the development of a sense of
community amongst members through a
process of identification between members.
The process of identifying with a
community enhances the individual’s
concern with collective processes and
group outcomes (Kramer, 1996, in
Nahapiet & Ghoshal, 1998), and has been
found to relate to expectations of
benevolent behaviour and community
participation (Chavis & Wandersman,
1990). Where the sense of community is
strong and benevolence is high, community
members are more likely to perceive
knowledge as a public good, owned and
maintained by the community. Wasko and
Faraj (2000) note:
“With a public good, the economically rational
action is to free-ride. [However,] the motivation to
exchange knowledge as a public good goes beyond
the maximisation of self-interest and personal gain.
People do not act only out of self-interest, but
forego the tendency to free-ride out of a sense of
fairness, public duty, and concern for their
community … People often behave altruistically
and pro-socially, contributing to the welfare of
others without apparent compensation” (Wasko &
Faraj, 2000: 161-2).
From the public good perspective, knowledge
sharing can be viewed as self-motivated
through a sense of moral obligation and a
general desire to be part of something
larger. Such pro-social behaviours lead to
the emergence of trust based on the
perceived benevolence of the community ,
whereby members expect that help will be
reciprocated should it be requested.
Conversely, if one’s sense of a
community’s
benevolence
is
low,
expectations of future reciprocity may
likewise be low, and knowledge sharing is
unlikely to be fostered. Furthermore, if
low perceptions of a community’s
benevolence are combined with high
perceptions
of
the
community’s
competence, this may exacerbate the fear
of losing face barrier discussed above .
Benevolence-based trust will contribute to
overcoming the fear of losing face by
creating the confidence that one will not
be criticised or made to look foolish when
engaging publicly in sharing one’s
knowledge.
Further to this, knowledge sharing can
be viewed from the perspective of BarTal’s (1976) theory of Generalised
Reciprocity. From this perspective, the
beneficiaries of knowledge-contributions
are likely to seek to reciprocate benevolent
actions with the collective, where direct
reciprocation is not possible. Hence, we
put forward the following proposition:
H2: One’s degree of trust in the benevolence of a
community is positively related to one’s
engagement in knowledge sharing with the
community.
Integrity is a much debated concept within the
trust literature. Sitkin and Roth (1993)
discuss how perceptions of integrity-based
trust are engendered within organisations
by the perception of congruence between
an individual’s values and the core
cultural values of the organisation; the
authors’ premise being that perceptions of
value incongruence will foster feelings of
distrust. This perspective sits close to
Mayer’s et al.’s definition of integritybased trust based on “perceptions that the
trustee adheres to a set of principles that
the trustor finds acceptable” (p. 719).
Mayer et al. elaborate on this
understanding by defining a number of
factors that influence the creation of
integrity-based trust, such as: the
independent verification of the trustee’s
integrity from reputable third parties;
perceptions that the trustee holds an
acceptable level of moral standards; and
demonstration of consistent behaviour
including congruence between a trustee’s
actions and words. The focus on the
alignment between an actor’s actions and
words is what Simon’s (2002) has defined
as behavioral integrity; he describes this
as the extent an individual is perceived to
“walk her talk”, adding that, conversely, it
30
reflects the extent to which she is seen to be
“talking her walk” (p. 19). Hence, trust in
the integrity of a virtual CoP might be
thought of as based in part on the
compatability of the community’s cultural
values with those of the trusting member,
the credibility of the community’s
reputation, and the consistency of
community members’ past behaviour such
as the extent to which actions are
congruent with words.
What can be derived from this
understanding of integrity is that such
perceptions are rooted in past behaviour.
Consistent and reliable past behaviour
creates confidence in future actions. If a
member expects that other members’
future behaviour may lack integrity, for
example, by acting dishonestly, unreliably
or in a manner that is otherwise
incongruent with her personal values, she
is not likely to readily engage in sharing
knowledge
with
the
community.
Conversely, she is likely to be more willing
to engage in cooperative interaction where
perceptions of honesty and expectations of
behavioural reliability are high. Hence,
H3: One’s degree of trust in the integrity of a
community is positively related to one’s
engagement in knowledge sharing with the
community.
Integrity-based
trust
H1
Competence-based
trust
H2
Online
Knowledge Sharing
H3
Benevolence-based
trust
Figure 2: The Antecedence of Trust to Knowledge Sharing
4
4.1
METHODOLOGY
Data Collection
For the fieldwork, CSC, a Fortune 500 global IT
services organisation provided access to a
suitable
virtual
community
of
practitioners. CSC employs over 79,000
people worldwide specializing in business
process
and
information
systems
outsourcing, systems integration and
consultancy services. The company have
been focussed on knowledge management,
and have been operating multiple online
‘knowledge communities’, for a number of
years.
Interviews with a number of community
leaders were held to develop an
understanding of the role of the
communities and the mechanisms used to
share knowledge. Access was granted to
survey the
organisation’s
Systems
Thinking Community, a global online
competence-based group of over 400
members that had been in existence for
over 4 years. The community’s main
purpose is to improve the organisation’s
business performance by applying the
tools of systems thinking. The community
develops decision simulation models by
running online systems thinking courses
via the Portal, with online workshops
using Lotus SameTime. Membership and
participation is entirely voluntary, and
when the survey was conducted, 120 of the
members were actively engaged in the
31
current course. Members received a link
to the survey sent out by email.
4.2
Measurement Development
4.2.1 Knowledge sharing
As knowledge sharing involves two or more
participants engaged in the provision and
acquisition of knowledge, it can therefore
be measured from the perspective of both
the source and the recipient of the
exchange.
However, this research
addresses the role of trust in the provision
of knowledge. Given that in the context of
online communities there may be multiple
recipients of shared knowledge, some of
whom the provider may be unaware of, a
logical approach to measuring knowledge
sharing would be to understand how to tap
into the construct from the source’s
perspective. Hence a number of metrics
were devised to measure the provision of
knowledge from the perspective of the
knowledge source3.
When measuring specific behaviour,
frequency of engagement in that behaviour
is often used as an indicant (for example,
Yoo et al. 2002). However, we argue that
in relation to knowledge sharing, such an
approach in itself is deficient. Knowledge
is intangible, therefore it cannot be easily
quantified. The frequency of engagement
in knowledge sharing behaviour does not
indicate the quality, usefulness or value of
the knowledge provided or acquired. For
example, a single contribution could have
more value than ten contributions
combined. Hence, measures that tap into
both the quality (A5-A8) and quantity (A1A2) of an individual’s provision of
knowledge were developed. Finally, the
degree to which an individual feels that
they engage in knowledge sharing will
provide an indication of the individual’s
knowledge sharing orientation. Hence,
knowledge sharing focus was measured as
a third dimension of knowledge sharing
behaviour; two measures were developed
to tap into this dimension (A3-A4).
3
See Appendix – Section A
4.2.2 Trust
McKnight et al. (2002) developed the Trusting
Beliefs Scale, which measures the degree
an individual believes another party to be
trustworthy. The original scales were
designed for examining levels of trust held
in online vendors and have been adapted
to fit the context of this study4. C1-C3
measure the degree to which one perceives
a community to be highly benevolent, C4C7 measure the degree to which one
perceives a community to behave with
high integrity, and C8-C11 measures the
trusting belief in the competence of the
community.
5
DATA ANALYSIS
After
partially
completed
and
spoiled
questionnaires were removed, this exercise
yielded 75 usable responses, representing
a response rate of 18%. This response
was appreciated given that only 27% of
the community members were actively
engaged
in
the
current
course.
Furthermore, Saunders et al. (2000) have
pointed out how time poor modern
organizations can be with regards to
research, and this is emphasized with
multi-national organizations such as this
study was based. The responses were
received from members primarily based
in the US (45%), UK (34%) and Australia
(11%). Members based in Switzerland,
Spain, Denmark and India made up the
remaining 10%. The respondents were
predominantly male (81%), with the
average length of tenure being 5 years, 10
months.
5.1
Validity and Reliability of Measures
By ensuring that the measurement scales are both
valid and reliable, we can increase the
credibility of the conclusions that can be
drawn by minimising the risk of
measurement error confounding the
results. The validity of a measurement
scale is “the degree to which it is free from
any systematic or non-random error”
(Hair et al., 1998: 90). In other words, the
4
See Appendix – Section C
32
extent to which the measures accurately
represent the construct that they are
intended to measure. Netemeyer et al.
(2003) elaborate on this understanding,
describing construct validity as “the degree
to which inferences legitimately can be
made from measures ... to the theoretical
constructs
on
which
those
operationalizations are based” (p. 71).
As the measures used for trust were
adapted from previously validated scales, it
was particularly important to assure the
validity and reliability of the newly
developed knowledge sharing scale.
Content validity provides a good
foundation upon which to establish the
validity of construct measurements
(Litwin, 1995); this was tested by asking
two of the community leaders and two
senior academics (the co-authors) for
feedback on the validity of the scales.
Following this feedback, a number of the
measures were reworded to aid clarity. A
negatively
loaded
question
was
incorporated into the survey as a reliability
check.
be evaluated via the loadings of each
indicator (Wasko & Faraj, 2005).
Each item loaded higher on the
expected construct than on other
constructs. The cross-dimension loadings
were high (.66 - .80) across the four
dimensions of our knowledge-sharing
construct, which is conceptually somewhat
unsurprising.
However, the crossdimension loadings were also high (.48 .70) across the three facets of trust, which
demonstrated that the three trust subconstructs are only mildly discriminant
from each other. In order to better
understand the relationship between the
theoretical
constructs,
a
principal
components analysis was conducted. As
we hypothesised a relationship between
the trust and knowledge sharing
constructs, an oblique rotation was used
(Pallant, 2005). Following theory and
specifying 1 knowledge management and 3
trust factors we have the results in Table
1.
Considering Table 1, the 3 items
generally cluster at their expected factor
grouping except the integrity item 6
T_INT which rather loads on the
benevolence factor.
The statement for
that item is “The community would keep
its commitments”. A follow up interview
is a possible means of discovering why the
respondents understood that question to
refer to community benevolence rather
than integrity. Meanwhile, the discussion
section has provided some explanation
why this could be the case.
5.1.1 Convergent and Discriminant
Validity
Convergent validity is a measure of the degree to
which two measures that are purported to
measure the same construct are correlated;
discriminant validity measures the degree
to which two conceptually similar
constructs are distinct (Hair et al., 1998).
Convergent and discriminant validity can
K-SHR7
K-SHR9
K-SHR6
K-SHR1
K-SHR2
K-SHR3
K-SHR4
K-SHR5
K-SHR8
11 T_CMP
8 T_CMP
1
0.92
0.91
0.90
0.88
0.82
0.82
0.81
0.79
0.79
2
3
4
0.90
0.89
33
10 T_CMP
0.86
9 T_CMP
0.82
1 T_BEN
0.85
2 T_BEN
0.73
6 T_INT
0.70
3 T_BEN
0.67
5 T_INT
0.38
4 T_INT
0.33
0.37
7 T_INT
0.40
Rotation Method: Oblimin with Kaiser
Normalization.
Rotation converged in 11
a
iterations.
0.45
0.57
0.50
0.47
Table 1: 4 Factor Extraction - Oblimin
6 DISCUSSION AND FUTURE RESEARCH
DIRECTIONS
All the three trust factors positively relate to
knowledge sharing in on-line communities
of practice thus upholding all the 3
hypotheses of this study. Each will now be
discussed.
6.1
Competent-based Trust
The positive hypothesis tests true that competentbased trust positively influences knowledge
sharing in virtual communities.
This
finding is in line with recent organisational
research that regards integrity (and
competence) as paramount to trust (Adler,
2001). As pointed out by Ardichvili et al
(2003, p 73) there was need to verify that
competence (and integrity) are major
-Fear of
losing face
Competence
Trust
components
of
trust
in
virtual
communities of practice. This verification
is what this finding of the study has done.
Of the three trust factors, competence is
the one that has all its original items
clearly cluster together. The more a
virtual
community
is
competent
(knowledgeable, capable and effective in
sharing knowledge), the more its members
will be inclined to share knowledge.
At least from the sample used, it can be
said that the passion for knowledge had a
greater effect than fear of losing face. It is
however suggested that a future study
takes a closer look at the moderating
influences of these two variables. Figure 3
outlines how their influences could be
conceptualised.
Knowledge
sharing
+Passion for
knowledge
Figure 3: Moderating influences of fear of losing face and passion for knowledge
34
perceived as only integrity but also as
competence and benevolence; hence, the
existence of some cross-loading on those
items.
Overall, we can conclude that integrity is
a component of trust which affects
knowledge sharing in on-line communities
of practice. This conclusion agrees with
earlier research that recognise integrity as a
trust component (cf Tschannen-Moran,
2000, pp 314, 318)
Fear of losing face should dampen the positive
influence of competence trust on
knowledge sharing whereas passion for
knowledge should do the opposite (cf
Ardichvili et al. 2003; Nahapiet and
Ghoshal 1998). For arguments supporting
the existence of these moderating variables,
please refer to section 3.
6.2
Benevolence-based Trust
Items C1 (community would act in my best
interest) and C3 (community is interested
in my well-being, not just its own) clearly
cluster together and support prove
hypothesis 2 that benevolent trust
positively affects the level of knowledge
sharing. When it comes to C2 (if required,
the community would do its best to help
me), the highest loading (0.73) is on the
benevolence factor but other loading (0.45)
is on the integrity factor. This suggests
that
respondents
perceived
the
community’s willingness to do the best to
help its members as not only benevolence
but also an integrity attribute.
On the whole, benevolence stands out as
a component of trust that affects
knowledge sharing in on-line communities
of practice. This finding is in consonant
with earlier research that include
benevolence as a component of trust (cf
Tschannen-Moran, 2000, pp 314, 318)
6.3
Integrity-based trust
Though they mostly cluster on the expected factor,
the integrity items are the most
problematic. In the first place, the fact
that the community would keep its
commitments (C6) is perceived by
respondents as benevolence instead of
integrity. This could be because, even
though they also acted as givers of
knowledge to each other, the respondents
regarded themselves more as students of
the community; taking knowledge from the
community to apply to their work situation
(see section 4.1).
Moreover, honesty and sincerity of the
community (C4, C5 and C7) are not
7 CONCLUSIONS
The research has upheld the three trust factors as
positively affecting knowledge sharing in
online communities. As discussed in 5.1.1
the cross loading of the factors suggest that
though they are theoretically distinct,
practically, they buttress each other to
influence knowledge sharing.
7.1
Implications for Knowledge Management
Professionals
Current KM research and practice has recognised
that the informal and trusting nature of
communities of practice is the bedrock of
knowledge sharing. The implication is that
KM practitioners need not be pro-active in
forming and running communities of
practice. Nonetheless, an understanding of
the trust components would guide
practitioners on how to create and support
knowledge sharing environment.
This
study has confirmed that trust is composed
of the three components of competence,
benevolence and integrity which are
positively related to knowledge sharing in
on-line communities of practice.
The
implication is that practitioners should
support the three values if they are to
encourage trust and knowledge sharing.
This research did not investigate how to
support the three values but a priori,
organisations should look for these three
qualities in their would-be recruits. Also,
KM practitioners can lead by example:
exhibit these qualities in their dealing with
other members of the organisation. They
35
can also take advantage of every
opportunity to encourage these attributes
in existing employees, who are potential
members of on-line communities of
practice.
The clearer-cut and most
significant component appears to be
competence but the findings show that the
factors enhance each other and therefore,
KM practitioners should be interested in
all the three components of trust.
ACKNOWLEDGMENTS
We would like to express our thanks to Computer
Sciences
Corporation
for
their
collaboration during this project, in
particular, for the company’s participation
in the empirical research.
C3
C4
C5
C6
C7
C8
C9
C10
APPENDIX. QUESTIONNAIRE
C11
SECTION A: KNOWLEDGE SHARING
A1 I frequently share my knowledge with
others in the community.
A2 I am one of the more active contributors
within the community.
A3 I make a conscious effort to spend time
engaged in activities that contribute
knowledge to the community.
A4 I try to share my knowledge with the
community.
A5 Other community members find my
knowledge-sharing contributions to be
useful.
A6 My contributions to the community enable
others to develop new knowledge.
A7 I am a knowledgeable contributor to the
virtual community.
A8 The knowledge I share with the community
has a positive impact on the business.
A9 Overall, I feel the frequency and quality of
my knowledge-sharing efforts are of great
value to the community.
SECTION C: TRUST
C1
I believe that the competency group’s
virtual community would act in my best
interest.
C2
If I required help, the community would
C12
do its best to help me.
The community is interested in my wellbeing, not just its own.
The community is truthful in its dealings
with me.
I would characterise the community as
honest.
The community would keep its
commitments.
The community is genuine and sincere.
The community is a competent and
effective source of expertise.
The community performs its role of
sharing knowledge very well.
Overall, the community is a capable and
proficient source of expertise and
knowledge.
In general, the community is very
knowledgeable.
I trust the community when I ask them
not to forward or share any CSC or client
sensitive material.
References
Adler, P. S. (2001). "Market, Hierarchy, and Trust: The
Knowledge Economy and the Future of
Capitalism." Organization Science, 12(2), 215-234.
Ardichvili, A., Page, V., and Wentling, T. (2002).
"Motivation and Barriers to Participation in Virtual
Knowledge-sharing Communities of Practice."
Reprint: <12 Reprint Status>
Ardichvili, A., Page, V., and Wentling, T. (2003).
"Motivation and barriers to participation in virtual
knowledge-sharing communities of practice."
Journal of Knowledge Management, 7(1), 64-77.
Argyris, C. and Schon, D. A. (1978). Organizational
Learning: A Theory of Action Perspective,
Addison-Wesley Publishing, Reading, MA.
Beaumont, P. and Hunter, L. C. (2002). "Managing
Knowledge Workers." CIPD, London.
Bhattacherjee, A. (2002). "Individual Trust in Online
Firms: Scale development and initial test." Journal
of Management Information, 19(1), 211-242.
Brown, J. S. and Duguid, P. (1991). "Organisational
Learning and Communities of Practice: Towards a
unified view of working, learning and innovation."
Organization Science, 2(1), 41-57.
36
Davenport, T. H. and Prusak, L. (1998). Working
Knowledge, Harvard Business School Press,
Boston, Massachusetts.
de Vaus, D. A. (1996). Surveys in Social Research,
Unwin Hyman, London.
Fisman, R. and Khanna, T. (1999). "Is trust a
historical residue? Information flows and trust
levels." Journal of Economic Behavior &
Organization, 38, 79-92.
Fukuyama, F. (1995). Trust: The social virtues & the
creation of prosperity, The Free Press, New York.
Gambetta, D. (1988). "Can we trust trust?" Trust:
Making and breaking cooperative relations, D.
Gambetta, ed., Basil Blackwell, New York, 213238.
Hair, J. S., Anderson, R. E., Tatham, R. L., and Black,
W. C. (1995). Multivariate Data Analysis,
Prentice-Hall, Engelwood Cliffs, NJ.
Herzberg, F. (1968). "One More Time: How Do You
Motivate
Employees?"
Harvard
Business
Review,(Special Issue - January 2003), 87-96.
Hildreth, P. and Kimble, C. (2000). "Communities of
Practice
in
the
distributed
international
environment."
Journal
of
Knowledge
Management, 4(1), 27-38.
Jarvenpaa, S. L., Knoll, K., and Leidner, D. E. (1998).
"Is Anybody Out There? Antecedents of Trust in
Global Virtual Teams." Journal of Management
Information Systems, 14(4), 29-64.
Jarvenpaa, S. L. and Staples, D. S. (2000). "The Use
of Collaborative Electronic Media for Information
Sharing: an exploratory study of determinants."
Journal of Strategic Information Systems, 9, 129154.
Kramer, R. M. and Tyler, T. R. (1996). Trust in
Organisations: Frontiers of theory and research,
Sage, Thousand Oaks, CA.
Lave, J. and Wenger, E. (1991). Situated Learning:
Legitimate Peripheral Participation, Cambridge
University Press, Cambridge, MA.
Lesser, E. and Everest, K. (2001). "Communities of
Practice: Making the most of intellectual capital."
Ivey Business Journal, 65(4).
Lesser, E. L. and Storck, J. (2001). "Communities of
Practice and Organisational Performance." IBM
Systems Journal, 40(4), 831-841.
Levin, D. Z., Cross, R., and Abrams, L. C. (2002a).
"The Strength of Weak Ties You Can Trust: The
mediating role of trust in effective knowledge
transfer." Rutgers University.
Marques de Sa, J. P. (2003). Applied Statistics using
SPSS, STATISTICA, and MATLAB, SpringerVerlag, Berlin.
Mayer, R. C., Davis, J. H., and Schoorman, F. D.
(1995). "An Integrative Model of Organisational
Trust." Academy of Management Review, 20(3),
709-734.
McKnight, D. H., Choudhury, V., and Kacmar, C.
(2002). "Developing and Validating Trust Measures
for e-Commerce: An Integrative Typology."
Information Systems Research, 13(3), 334-359.
McKnight, D. H., Cummings, L. L., and Chervany, N.
L. (1998). "Initial Trust Formation in New
Organizational Relationships." Academy of
Management Review, 23(3), 473(18).
McMillan, D. W. and Chavis, D. M. (1986). "Sense of
Community: A Definition and Theory." Journal of
Community Psychology, 14(1), 6-23.
Mishra, A. K. (1996). "Organisational Responses to
Crisis: The Centrality of Trust." R. M. K. &. T. R.
Tyler, ed., Sage, Thousand Oaks, CA, 261-287.
Nahapiet, J. and Ghoshal, S. (1998). "Social Capital,
Intellectual Capital and the Organizational
Advantage." Academy of Management Review,
23(2), 242-266.
Preece, J. (2000). Online Communities: Designing
Usability, Supporting Sociability, John Wiley &
Sons, Chichester, England.
Rheingold, H. (1993). The Virtual Community:
Homesteading on the Electronic Frontier, AddisonWesley, Reading, MA.
Ring, P. S. and Van de Ven, A. H. (1994).
"Developmental
Processes
of
Cooperative
Interorganizational relationships." Academy of
Management Review, 19, 90-118.
Ruggles, R. L. (1998). "The State of the Notion:
knowledge management in practice." California
Management Review, 40(3), 80-89.
Saunders, M., Lewis, P., and Thornhill, A. (2000).
Research Methods for Business Students, Pearson
Education, Harlow.
Scarbrough, H. (1999). Knowledge Management: A
literature review, Institute of Personnel and
Development.
Scholl, W., Konig, C., Meyer, B., and Heisig, P.
(2004). "The Future of Knowledge Management: an
international delphi study." Journal of Knowledge
Management, 8(2), 19-35.
Simons, T. (2002). "Behavioural Integrity: The
Perceived Alignment Between Managers' Words
37
and Deeds as a Research Focus." Organization
Science, 13(1), 18-35.
Tsai, W. and Ghoshal, S. (1998). "Social Capital and
Value Creation: The Role of Intrafirm Networks."
Academy of Management Journal, 41(4), 464-476.
Tschannen-Moran (2001) “Collaboration and the need
for trust” Journal of Educational Administration
39(4), 308-31.
Van Beveren, J. (2002). "A Model of Knowledge
Acquisition
that
Refocuses
Knowledge
Management."
Journal
of
Knowledge
Management, 6(1), 18-22.
Wang, C. L. and Ahmed, P. K. (2003). "Structure and
structural dimensions for
knowledge-based
organizations." Measuring Business Excellence,
7(1), 51-62.
Wasko, M. M. and Faraj, S. (2000). ""It is what one
does": why people participate and help others in
electronic communities of practice." Journal of
Strategic Information Systems, 9(2-3), 155-173.
Wenger, E. (1998). Communities of Practice:
Learning, meaning and identity, Cambridge
University Press, Cambridge.
Wenger, E., McDermott, R., and Snyder, W. (2002).
Cultivating Communities of Practice: A guide to
managing knowledge, Harvard Business School
Press.
Wenger, E. and Snyder, W. M. (2000). "Communities
of Practice: The organisational frontier." Harvard
Business Review,(Jan-Feb 2000), 139-145.
Yoo, W.-S., Suh, K.-S., and Lee, M.-B. (2002).
"Exploring
Factors
Enhancing
Member
Participation in Virtual Communities." Journal of
Global Information Management, 10(3), 55-71.
Dr Abel Usoro is a Lecturer in the School of
Computing, Mark W Sharratt was his MSc student,
and Eric Tsui is a Professor at the HK Polytechnic
University in Hong Kong. Prof Tsui facilitated access
to the sampling frame and also collaborates in the
research.
38
Revise
New
Reuse
Proposed
Retrieve
Retain
Problem
proposed
case
Case
Base
Similar
information
experience
Situation
Solution
solution
Cases
39