Semantics and Analysis of DMN Decision Tables

Diego Calvanese, Marlon Dumas, Ülari Laurson, Fabrizio Maria Maggi, Irene Teinemaa

University of Tartu, Estonia | Free University of Bozen-Bolzano, Italy

Semantics and Analysis of

DMN Decision Tables

BPM’2016, Rio de Janeiro, 21 September 2016

2

• http://dmn.cs.ut.ee

Demo

3

Geometric interpretation

4

Line sweeping

5

Finding overlapping rules

(A, C)

6

Finding overlapping rules

E(A, C)(A,E)(C,E)

F

(C,F)

7

A C

E F

{A, C, E}{C, E}

Signavio{A, C} {A, E}, {C,E}{C, E}

Grouping rules:

Maximum clique decomposition

8

Finding missing rules

(A, C)

9

Finding missing rules

(A, C)

10

• LendingClub credit rating dataset• 23K loans rated in categories A, B, ... G

• Loan collection Decision tree Decision table• 3, 5 and 7 columns

• 500, 1K, 5K rules

• Mixture of categorical and numerical columns

• Noise injection• 10% of rows randomly chosen, one column enhanced (or shrank for missing rules)

• Baseline: Signavio DMN editor

Experimental evaluation

11

3 columns 5 columns

#Rules 499 998 1492 505 1000 1506

#Overlapping rule sets

Our approach 239 725 1 204 172 363 632

Signavio 1 226 10 920 23 115 679 3 692 8 921

#Missing rules

Our approach 117 330 726 136 254 462

Signavio 668 2 655 5 386 563 2 022 4 832

Experimental Results# of overlapping rules and missing rule-sets reported

Execution times (1500 rules x 5 columns)• Overlapping rules: < 3 seconds (with optimizations not reported in the paper)• Missing rules: < 2 seconds

12

Epilogue: Decision table simplification

13

Epilogue: Decision table simplification

14

3 columns 5 columns

#Columns modified 1 2 3 1 2 3

Our approach 104 107 96 105 104 119

Pollack approach 178 178 158 165 163 206

All combination 97 101 90 102 100 110

# of rules after simplification (100 rules)

15

• Formalized decision tables and two analysis problems

• Proposed algorithms for:• Detecting overlapping rules

• Detecting missing rules• Simplify a decision table via rule merging (not in this paper)

• Scalable and more compact feedback than state-of-the-art tools

• Algorithm implemented in dmn-js • Online editor: http://dmn.cs.ut.ee

• Source code: https://github.com/ulaurson/dmn-js

• Algorithm for decision table optimization: http://goo.gl/6QASgc

Conclusion