Stochastic Process Discovery by Weight Estimation

Abstract

Many algorithms now exist for discovering process models from event logs. These models usually describe a control flow and are intended for use by people in analysing and improving real-world organizational processes. The relative likelihood of choices made while following a process (i.e., its stochastic behaviour) is highly relevant information which few existing algorithms make available in their automatically discovered models. This can be addressed by automatically discovered stochastic process models.We introduce a framework for automatic discovery of stochastic process models, given a control-flow model and an event log. The framework introduces an estimator which takes a Petri net model and an event log as input, and outputs a Generalized Stochastic Petri net. We apply the framework, adding six new weight estimators, and a method for their evaluation. The algorithms have been implemented in the opensource process mining framework ProM. Using stochastic conformance measures, the resulting models have comparable conformance to existing approaches and are shown to be calculated more efficiently.

Online Incremental Mining Based on Trusted Behavior Interval

“…Since the direct successor of logs reflects the relationship between traces in terms of behavioral structure [31], the direct successor distance is defined on its basis to measure the difference in behavior of two traces. Definition 9: Direct successor distance.…”

Probabilistic Trace Alignment

“…This approach favors traces providing optimal alignment cost: the resulting ranking cannot be used to provide a trade-off between trace probability and alignment cost. As the model has non-stochastic Petri Net as a reference model, the computation of the aforementioned probability for activities' mismatch requires a log file for providing such an estimation as the more recent work [8], while our proposed solution might as well estimate trace probability by directly loading Stochastic Petri Nets. We deduce that our proposed solution proves to be more general that this first attempt to probabilistic trace alignments.…”

“…t ∈E(M ) W (t ) [8]. Notice that, by construction, the probabilities associated to all enabled transitions in a marking always add up to 1.…”

“…We used the training set to generate either an SWN, using the approach presented in [18], or a BPMN with only exclusive gates using Split Miner 2.0 [20]. Our implementation approxProbTraceAlign 3 takes as an input both models, while internally converting the latter into a Petri net [21], which is later on converted into an SWN via a firing weight estimator: we chose the Fork and the PairScale estimators from [8] and we denote as Constant a naive estimator assuming that all the transition enabled in a given marking are equiprobable. The user can select these estimators as well as the others from [8] from our GUI.…”

“…Our implementation approxProbTraceAlign 3 takes as an input both models, while internally converting the latter into a Petri net [21], which is later on converted into an SWN via a firing weight estimator: we chose the Fork and the PairScale estimators from [8] and we denote as Constant a naive estimator assuming that all the transition enabled in a given marking are equiprobable. The user can select these estimators as well as the others from [8] from our GUI. Finally, we loaded the testing set, which contains the log traces to be aligned against the previously loaded model (such traces can be potentially filtered in the tool).…”