Marlon Dumas University of Tartu, Estonia Estonian Theory Days | - PowerPoint PPT Presentation

Marlon Dumas University of Tartu, Estonia Estonian Theory Days | 3-4 October 2015

Process Mining Discovery discovered ¡model Deviance Performance event ¡log Difference Enhanced ¡model diagnos7cs event ¡log’ ü / û Conformance 2 input ¡model

Automated Process Discovery CID ¡ Task ¡ Time ¡Stamp ¡ … ¡ 13219 ¡Enter ¡Loan ¡Applica3on ¡ 2007-­‑11-­‑09 ¡T ¡11:20:10 ¡ -­‑ ¡ 13219 ¡Retrieve ¡Applicant ¡Data ¡ 2007-­‑11-­‑09 ¡T ¡11:22:15 ¡ -­‑ ¡ 13220 ¡Enter ¡Loan ¡Applica3on ¡ 2007-­‑11-­‑09 ¡T ¡11:22:40 ¡ -­‑ ¡ 13219 ¡Compute ¡Installments ¡ 2007-­‑11-­‑09 ¡T ¡11:22:45 ¡ -­‑ ¡ 13219 ¡No3fy ¡Eligibility ¡ 2007-­‑11-­‑09 ¡T ¡11:23:00 ¡ -­‑ ¡ 13219 ¡Approve ¡Simple ¡Applica3on ¡ 2007-­‑11-­‑09 ¡T ¡11:24:30 ¡ -­‑ ¡ 13220 ¡Compute ¡Installements ¡ 2007-­‑11-­‑09 ¡T ¡11:24:35 ¡ -­‑ ¡ … ¡ … ¡ … ¡ … ¡ Notify Retrieve Rejection Applicant Data Enter Loan Application Approve Simple Compute Application Installments Notify Eligibility Approve Complex 3 Application

Automated Process Discovery • Relations-based – Alpha 4

Alpha Algorithm A A B C D C A C B E F B D A B C E G • Direct successors: • Causality: A > B, B > C, C > D, A → B, C → D, A → C, B → E, A > C, C > B, B > E, E > F C → E, E → F, E → G , B → D C > E, E > G • Concurrency: B > D B ║ C • Exclusiveness: all other pairs 5

Alpha Relations Matrix A B C D E F G # → → → # # # A B ← # || → → # # C ← || # → → # # # ← ← # # # # D # ← ← # # → → E # # # # ← # # F # # # # ← # # G 6

Alpha Algorithm – Patterns a → b, ⇔ a → c, b ║ c A B C D E F G A # → → # # # # B ← # || → → # # C ← || # → → # # D # ← ← # # # # E # ← ← # # → → F # # # # ← # # G # # # # ← # # 7

Automated Process Discovery • Relations-based – Alpha: lossy (Badouel 2012) – Alpha++, Alpha#, Alpha$ – Heuristics miner (frequency information) • Genetic • Region theory • Petri net synthesis • Integer Linear Programming (ILP) • … 8

Automated Process Discovery Simplicity minimal size & structural complexity Automated Generalization Precision process parses traces of the does not parse discovery process not included traces not in the log method in the log Fitness parses the traces of the log 9

Conformance Checking ? � 10

Alignment-Based Conformance Check Log Model Alignment A B C D E A B B C E Fitness Precision How much behavior of the log How accurate is the model is captured by the model? describing the log? 11 Munoz-Gama et al. Petri nets 2013

Imprecision of Alignment-Based Conformance Checking • {ABCD, ACBD} à 100% • {ABD, ACBCD} à 100% • {ACD, ABCBD} à 100% 12

Deviance Mining T 1 ¡<e 11 [d 111 :v 111 , ¡…, ¡d 11n :v 11n ] ¡e 12 [d 121 :v 121 , ¡…, ¡d 12m :v 12m ] ¡… ¡e 1p [d 1p1 :v 1p1 , ¡…, ¡d 1pm :v 1pm ]> ¡ … ¡ T q ¡<e q1 [d q11 :v q11 , ¡…, ¡d q1n :v q1n ] ¡e q2 [d q21 :v q21 , ¡…, ¡d q2m :v q2m ] ¡… ¡e qp [d qp1 :v qp1 , ¡…, ¡d qpm :v qpm ]> ¡ ¡ Find a function F: Trace à Boolean (or probability [0…1]) s.t. • F is an accurate approximation of the given labeling • F is explainable, e.g. set of simple predicates 13

Deviance Mining via Sequence Classification • Apply discriminative sequence mining methods to extract features characteristic of one class • Build classification models (e.g. decision trees) • Extract difference diagnostics from classification model 14 C. Sun et al. Mining explicit rules for software process evaluation. ICSSP’2013.

No Unified Foundation Automated process discovery • Behavioral relations, theory of regions, ILP, … Conformance checking • Replay, alignments Deviance mining • Model delta analysis, sequence classification ≠ 15

(Prime) Event Structures • Model of concurrency based on events (occurrences of actions) and three relations – Causality – Conflict – Concurrency C E B C A E A B D D E 16

τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ Petri Nets à à Event Structures a b c d c b d 17 d d

Nets With Cycles à à Prefix Unfolding Petri net N Causality-preservng Complete prefix prefix unfolding unfolding 18

Comparison of Event Structures Partially {},{} ES1 Synchronized B C match(A Product (PSP) ) A E {A},{A} match(B D ) In ES1, tasks C and B are {A,B},{A,B} mutually exclusive, while ? � match(D in ES2, B precedes C � ) {A,B,D},{A,B,D} C E ES2 left_hide( C) A B {A,B,C,D},{A,B,D} match(D ) D E {A,B,C,D,E}, {A,B,D,E} 19 Armas-Cervantes et al. Behavioral Comparison of Process Models Based on […] Event Structures. BPM’2014

Event Structures for Process Mining Deviance mining Conformance checking Process discovery 20

Event Logs è è Event Structures Run Merger Concurrency Oracle 5 2 3 B || C 21

Event Structures for Log Delta Analysis 22 van Beest et al. Log delta analysis: Interpretable differencing of business process event logs. BPM’2015

Event Structures for Log Delta Analysis In L1, task C can be skipped after B, whereas in L2 it cannot � 23 van Beest et al. Log delta analysis: Interpretable differencing of business process event logs. BPM’2015

Log Delta Analysis vs. Sequence Classification Sequence classification 106-130 statements IF |“NursingProgressNotes”| > 7 .5 THEN L1 � IF |“Nursing Progress Notes”| ≤ 7 .5 AND |“Nursing Assessment”| > 1.5 � THEN L2 � … � Log delta analysis 48 statements In L1, “Nursing Primary Assessment” is repeated after “Medical Assign Start” and “Triage 448 cases 363 cases, Request”, while in L2 it is not. � 7329 events 7496 events … � 24 van Beest et al. Log delta analysis: Interpretable differencing of business process event logs. BPM’2015

Event Structures for Conformance Checking B C Check credit Assess loan B C A history risk E Receive Assess A E application D eligibility Appraise D property B Check credit ABDE history C E C ADBE A E Assess loan A B ACDE risk Receive Assess application eligibility ADCE Appraise property D E D 25

Event Structures for Process Discovery? ABDE ACDE ACDF Merge Synth Fold . 26

Process Mining Reloaded 27