Abductive Workflow Mining Using Binary
Resolution on Task Successor Rules
Scott Buffett
National Research Council Canada
University of New Brunswick
RuleML 2008
Orlando, Florida, October 30, 2008
Workflow Mining
• Transaction log containing a number of events
• Each event is labeled by a task and a case
• Tasks executed a case give a trace
• ABCE, ABDE, ACBE, ADBE
B
A
C
D
E
Task Case
A
1
A
2
A
3
C
3
B
1
B
2
A
4
B
3
D
2
C
1
E
1
B
4
D
4
E
2
E
3
E
4
2
Using Workflow to
Measure Compliance
• Compare observed activity with accepted model
B
• ACDE?
A
C
E
D
• If discrepancies exist, behavior may be non-compliant
3
Problems with Using
Workflow
• Detected non-compliant behaviour does not imply inappropriate
activity
• Behaviour might be OK, but not captured during workflow mining
• Workflow model not 100% accurate
• Errors in task / case labelings
• Noise
• Process may have evolved or changed
4
Solution
• Identify the tasks that are of high importance
• Example process: Bank loan application
–
–
–
–
–
–
A: Enter financial data
B: Access credit report
C: Process loan application form
D: Process pre-approved loan form
E: Approve loan
F: Reject loan application
Task B: Access
credit report
D
• Cases observed:
ABCE, ABCF, ADE
E
A
B
C
F
5
Abductive Workflow
Mining
• Reduce the problem to mining workflow that necessarily
implies that the critical activity must be executed
• We call this “abductive workflow”
D
Presence of task C
(process loan application
form) implies B
E
A
B
C
F
6
Some Example Abductive
Workflows for Critical
Task “B”
C
D
A
F
C
E
B
F
C
F
C
F
A
B
A
B
C
7
Desirable Properties
• Trace minimality:
C
A
B
C
• Completeness:
C
C
F
8
Finding Desirable
Workflows
• Task successor rules
• Indicate activity that immediately follows certain tasks
• Example workflow traces:
– PQR, PRS, RMN, TVQ
• Critical activity: R
9
Divide Positive and
Negative Traces
Traces: PQR, PRS, RMN, TVQ, Critical: R
• Positive traces:
PQR, PRS, RMN
• Negative traces:
TVQ
10
Remove Critical
Activity
Traces: PQR, PRS, RMN, TVQ, Critical: R
• Positive traces:
• Remove critical activity:
PQR, PRS, RMN
PQ, PS, MN
• Negative traces:
TVQ
11
Add Dummy Tasks
Traces: PQR, PRS, RMN, TVQ, Critical: R
• Positive traces:
• Remove critical activity:
• Add dummy tasks:
PQR, PRS, RMN
PQ, PS, MN
PQw’, PSw’, MNw’
• Negative traces:
• Add dummy tasks:
TVQ
TVQw0’
12
Task Successor Rules
• One for every subsentence in positive traces (except w’)
• Positive: PQw’, PSw’, MNw’
• Negative: TVQw0’
• Rules:
P -> Q, S
S -> w’
N -> w’
PS -> w’
Q -> w’, w0’
M -> N
PQ -> w’
MN -> w’
13
Finding Abductive
Workflows
• Convert to CNF
~P, Q, S
~S, w’
~N, w’
~P, ~S, w’
~Q, w’, w0’
~M, N
~P, ~Q, w’
~M, ~N, w’
• Binary resolution, generate clauses where w’ is the only
positive literal
~PQS
~P~Qw’
~PSw’
~Sw’
P -> w’
~Pw’
14
Finishing the Example
• Complete set of abductive traces:
– P, S, M, N, PS, MN, PQ
• Task-minimal abductive traces:
– P, S, M, N
PQR, PRS, RMN
TVQ
• Complete workflows:
– {P,M}, {P,N}, {P,M,N}, {P,S,M}, {P,S,N}, {P,S,M,N}
• Complete, trace-minimal:
– {P,M}, {P,N}
15
Results
• Test size reduction of abductive workflows
• Uses a naïve method for finding abductive workflows, complete
but not necessarily minimal
• Thus provides a lower bound on size reduction
• Mines entire workflow and extracts abductive workflow
• Ran on example log files accompanying ProM software
Our Miner
Alpha Miner
Statistic (Avg)
Orig. Workflow
Abd. Workflow
Orig. Workflow
Abd. Workflow
# of transitions
156.2
37.2
12.0
5.6
# of arcs
318.6
77.0
33.2
15.6
16
Conclusions
• Abductive workflows provide a condensed model, adequate for
validating particular critical activity
• Mitigate a number of problems inherent in compliance checking
• Rules can help determine such workflows, with desirable
properties
• Potential for significant decreases in size of workflow model was
demonstrated
17
18
© Copyright 2025 Paperzz