Stochastic conformance checking based on variable-length Markov chains

Conformance checking is central in process mining (PM). It studies deviations of logs from reference processes. Originally, the proposed approaches did not focus on stochastic aspects of the underlying process, and gave qualitative models as output. Recently, these have been extended in approaches for stochastic conformance checking (SCC), giving quantitative models as output. A different community, namely the software performance engineering (PE) one, interested in the synthesis of stochastic processes since decades, has developed independently techniques to synthesize Markov Chains (MC) that describe the stochastic process underlying program runs. However, these were never applied to SCC problems. We propose a novel approach to SCC based on PE results for the synthesis of stochastic processes. Thanks to a rich experimental evaluation, we show that it outperforms the state-of-the-art. In doing so, we further bridge PE and PM, fostering cross-fertilization. We use techniques for the synthesis of Variable-length MC (VLMC), higher-order MC able to compactly encode complex path dependencies in the control-flow. VLMCs are equipped with a notion of likelihood that a trace belongs to a model. We use it to perform SCC of a log against a model. We establish the degree of conformance by equipping VLMCs with uEMSC, a standard conformance measure in the SCC literature. We compare with 18 SCC techniques from the PM literature, using 11 benchmark datasets from the PM community. We outperform all approaches in 10 out of 11 datasets, i.e., we get uEMSC values closer to 1 for logs conforming to a model. Furthermore, we show that VLMC are efficient, as they handled all considered datasets in a few seconds.