Handling Duplicated Tasks in Process Discovery by Refining Event Labels

Xixi Lu (x.lu@tue.nl)1

Dirk Fahland1

Frank van den Biggelaar2

Wil van der Aalst1

1Eindhoven University of Technology2Maastricht University Medical Center 1

Overview

• (5 min) Research Problem• (10 min) Approach• (5 min) Evaluation and Conclusion

2

VA CB A

Duplicated Tasks –Events Having Imprecise Labels

3

CA

BS+

X X

A B ACV S

C B SAC A B

VV S

Discover VABC

S

Record

+

4

A B ACV S

C B SAC A B

VV S

Input:Imprecise log

Research Problem

Research Problem

5

So that1. all traces & events preserved (no filtering)2. model is more precise (than without refining)3. we can explore different refinement of labels interactively

(because we don’t know correct label refinement from given log)

Discover(IM, ILP, …)

Model

Refine labels of events

Input:Imprecise log

Output:Refined log

Refining Event Labels

6

imprecise Precise

DiscoverDiscover Discover

C B AC A B

A B AC C2 B2 A2

C2 A2 B2

A1 B1 A3C1

C2 B2 A2

C3 A4 B3

A1 B1 A3C1

C2 B2A2

A1 B1 A3C1

C2 B2 A2

C3 A4 B3

A1 B1 A3C1

C BA

A B AC

Discover

C2 B2 A2

C3 A4 B3

A1 B1 A3C1Refining Label

ABC

Refining Event Labels

7

Precise

Discover Discover

C B AC A B

A B AC C2 B2 A2

C2 A2 B2

A1 B1 A3C1

C2 B2 A2

C3 A4 B3

A1 B1 A3C1

C2 B2A2

A1 B1 A3C1

C2 B2 A2

C3 A4 B3

A1 B1 A3C1

C BA

A B ACSystem unknown!

We don’t know what is optimal!

Discover

C2 B2 A2

C3 A4 B3

A1 B1 A3C1Refining Label

ABC

imprecise

Discover

Related Work

8

Discover(IM, ILP, …)

Process Model

Refine Labelsof Events

Imprecise log Refined log

(c) Cluster traces

(b) New Discovery algorithm

(a) Relabelmodel elements

(d) Filter

Filtered log

Clusters …

1. Detectimprecise

labels3. Refine

horizontally4. Refinevertically

Imprecise label candidates

Horizontal clusters of events

Vertical clusters of events

Approach

9

Refined log

2. Mapsimilar events

{x, …}

Mappings

Imprecise log

a) Manual by user inputb) Automated detection

1. Detectimprecise

labels3. Refine

horizontally4. Refinevertically

Imprecise label candidates

Horizontal clusters of events

Vertical clusters of events

Approach

10

Refined log

2. Mapsimilar events

{x, …}

Mappings

Imprecise log

Mapping Between Events

11

Trace 1

Trace 2

Mapping

AV SA CB

SV C AB

Quantify Similarity of Mapped Events

12

… based on “structural context of events”

AV SA CB

SV C AB

Quantify Similarity of Mapped Events

13

… based on “structural context of events”(1) Differences in neighbors

Different Same

Cost = 2

AV SA CB

SV C AB

Quantify Similarity of Mapped Events

14

… based on “structural context of events”(1) Differences in neighbors(2) Differences in structure

Distance = 4

Distance = 3

Cost = 2+1

AV SA CB

SV C AB

Quantify Similarity of Mapped Events

15

… based on “structural context of events”(1) Differences in neighbors(2) Differences in structure

Distance = 2

Distance = 1

Cost = 2+1+1+1+0

AV SA CB

SV C AB

Quantify Similarity of Mapped Events

16

… based on “structural context of events”(1) Differences in neighbors(2) Differences in structure

Cost = 2+3Cost = 4+6

Cost = 4+4 Cost = 0+3

Cost = 2+4 Cost so far = 6+10+8+5+3

AV SA CB

SV C AB

Quantify Similarity of Mapped Events

17

… based on “structural context of events”(1) Differences in neighbors(2) Differences in structure(3) #Dissimilar events

Xixi Lu, Dirk Fahland, Frank J.H.M. van den Biggelaar, and Wil M.P. van der Aalst. Detecting Deviating Behavior without Models. In BPM Workshops 2015, volume 256, pp.126-139, Springer International Publishing, 2015.

There is an algorithm to compute an optimal mapping :

AS SA CB

SS C AB

Total Cost = 6+10+8+5+3 +1

1. Detectimprecise

labels3. Refine

horizontally4. Refinevertically

Imprecise label candidates

Horizontal clusters of events

Vertical clusters of events

Approach

18

Refined log

2. Mapsimilar events

{x, …}

Mappings

Imprecise log

3. Refine horizontally using cost

19

AV SA CB

SV C AB

3. Refine horizontally using cost

20

AV SA CB

SV C AB

Trace 1

Trace 2

SV C BATrace 3

20

V SA CBTrace 0

10 10 10

Imprecise label candidates

20 20

15 15 15

a) Normalize costs w.r.t. maximal cost seen in the log

3. Refine horizontally using cost

21

1

0.5 0.5 0.5

Imprecise label candidates

1 1

0.75 0.75 0.75

a) Normalize costs w.r.t. maximal cost seen in the log

b) Set variant threshold,say, 0.8

c) Remove edges if cost > variant threshold

AV SA CB

SV C AB

SV C BA

V SA CB

Trace 1

Trace 2

Trace 3

Trace 0

3. Refine horizontally using cost

22

0.5 0.5 0.5

0.75 0.75 0.75

a) Normalize costs w.r.t. maximal cost seen in the log

b) Set variant thresholdsay, 0.8

c) Remove edges if cost > variant threshold

A1V SA1 C1B1

SV C2 A2B2

SV C2 B2A2

V SA1 C1B1

Trace 1

Trace 2

Trace 3

Trace 0

4. Refine vertically using frequency

23

I) Set unfolding threshold,say, 60%

II) Refine labels if freq > unfolding threshold

Is this A1 a 2nd iteration of a loop?

A1V SA1 C1B1

V SA1 C1B1

Trace 1

Trace 0

2 1

4. Refine vertically using frequency

24

I) Set unfolding threshold,say, 40%

II) Refine labels if freq > unfolding threshold

1

Or is this a different task?

A3V SA1 C1B1

V SA1 C1B1

Trace 1

Trace 0

2

1. Detectimprecise

labels3. Refine

horizontally4. Refinevertically

Imprecise label candidates

Approach

25

Refined log

2. Mapsimilar events

{x, …}

Mappings

Imprecise log

Variant threshold

Unfolding threshold

Iterative

Implemented in ProM

• There will be a demo in the demo track.

26

Detect imprecise labels

Refine horizontally

Refine vertically

• 3 experiments, each 600 models, k transitions w/ same labelE1) Default parameters, k = 4E2) Adaptive parameters, k = 4E3) 1 in loop, k = 2

Evaluation

27

Refining Label

Discover

Discover

Generate

Improved?Imprecise log Refined log

Compute qualities

Experiment Result - Example

28

System precision improved by 0.68System recall is 1

Precision w.r.t. log improved by 0.55

B

B B B B

𝑡𝑡1𝑡𝑡2

𝑡𝑡3 𝑡𝑡4B

BB B

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

E2) Adaptive

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

E1) Default

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

E3) 1 in loop

Experiment Result -F1-score of and w.r.t.

29

Improved 35% ~ 46% Models Improved 89% Models Improved 60% Models

Experiment Results and Limitations

30

Real-life Example : Hospital Data

31

1st Iteration : refined “surgery” events

2nd & 3rd Iterations: refine “consultation” events

Conclusion and Future Work

• Improved logs (and process discovery) by refining labels for up to 87%• Interactive and explorative

• Future work• Different ways to compute similar events using context of events• Log preprocessing framework

• Integrated in “Log to Model Explorer”• Supporting clustering, filtering and label refining

• Come to the Demo!

32

Questions?

33