Relevance Feedback Retrieval
of Time Series Data
Eamonn J. Keogh & Michael J. Pazzani
Prepared By/
Fahad Al-jutaily
Supervisor/
Dr. Mourad Ykhlef
IS531
Sunday, 16 April 2006
INTRODUCTION(1/2)
• Time series account for a significant portion
of the data stored in business, medical,
engineering, and social science databases.
• There are innumerable statistical tests one
can perform on time series, such as
determining autocorrelation coefficients,
measuring linear trends, etc.
INTRODUCTION(2/2)
• Much of the utility of collecting this data,
comes from the ability of humans to
visualize the shape of the (suitably plotted)
data. For example:
– Cardiologists view ECGs to diagnose arrhythmias.
– Chartists examine stock market data, searching for
certain shapes that are thought to be indicative of a
stock’s future performance.
– Astronomers examine star light curves (the changes in
frequency over time) to classify stars.
Issues in time series retrieval(1/2)
•
Two important issues in time series retrieval
have not yet been explored.
1. Relevance Feedback: In time series domains, as
in text domains, users may not initially know
how to form a query to retrieve precisely what
they are looking for.
• The paper shows, the relevance feedback may be
applied to retrieval of time series data to learn
which sections of the time series are most
significant.
Issues in time series retrieval(2/2)
•
Subjectivity of similarity: Most work on
retrieval of time series has used Euclidean
distance as the distance measure.
However, there is little evidence that
Euclidean distance maps onto human
intuition of similarity.
Example: Clusters of four small data sets where
similarity is determined by Euclidean distance.
Time Series Similarity Solution
• The “correct” distance measure depends
upon the user and problem, and it should be
continuously learned as the user interacts
with the system.
REPRESENTING TIME
SERIES(1/2)
• Using the original ‘raw’ data for query by
content in time series databases is
computationally expensive (especially for
an interactive system) and fails to abstract
key features of the data.
• What is needed?!!! is a higher-level
representation.
REPRESENTING TIME
SERIES(2/2)
• Piece-wise linear segmentation, which attempts to
model the data as sequences of straight lines.
• It provides a useful form of data compression and
noise filtering.
Notation(1/2)
• A time series, sampled at k points, is
represented as an uppercase letter such as A.
• The segmented version of A, containing K
linear segments, is denoted as a bold
uppercase letter such as A.
– where A is a 5-tuple of vectors of length K.
– A => {AXL, AXR, AYL, AYR, AW}
Notation(2/2)
• The ith segment of sequence A is
represented by the line between (AXLi
AYLi) and (AXRi AYRi), and AWi, which
represents the segment's weight.
Comparing Time Series(1/3)
• To retrieve time series, we need a distance
(or conversely similarity) measure to
compare the query with segments stored in
the database.
• Adopting a distance measure, DS, that
approximates the Euclidean distance
measure on the ‘raw’ data.
Comparing Time Series(2/3)
• It is convenient for notational purposes to assume
that the endpoints of the two sequences being
compared are aligned in the X-axis.
Comparing Time Series(3/3)
• The overall distance between two sequences is
simply the square root of the summation over all
such slices.
• The weighted distance between two
sequences of length i is:
Merging Time Series(1/2)
• The basic idea is that the merge operator
takes two sequences as input and returns a
single sequence whose shape is a
compromise between the two original
sequences.
Merging Time Series (2/2)
QUERY REFINEMENT VIA
RELEVANCE FEEDBACK
• Relevance feedback is the reformulation of
a search query in response to feedback
provided by the user for the results of
previous versions of the query.
Query
Output
Feedback
Formulating a New Query(1/2)
1. An initial query sequence Q is used to rank all sequences
in the database. This query may be hand drawn by the user
or it may be a sequence or subsequence from the database.
2. The best n sequences are retrieved and shown to the user.
3. The user assigns ratings to the retrieved sequences on a
scale from –3 to +3. After
4. the user has evaluated the top n sequences, the query
update rule that is used to produce a new query Qnew, and
5. The search process begins again.
Formulating a New Query(2/2)
SUBJECTIVITY OF SIMILARITY(1/2)
• It is still possible that a query using this
representation could fail to retrieve an item
that the user would have found relevant.
• Consider a problem of offset translation,
Two similar (or even identical) shapes can
have an arbitrarily large dissimilarity
because they are separated in the Yaxis.
SUBJECTIVITY OF SIMILARITY(2/2)