Refinement Mining: Using Data to Sift Plausible Models

Abstract

Process mining techniques have been developed in the ambit of business process management to extract information from event logs consisting of activities and then produce a graphical representation of the process control flow, detect relations between components involved in the process and infer data dependencies between process activities. These process characterisations allow the analyst to discover an annotated visual representation of the conceptual model or the performance model of the process, check conformance with an a priori model to detect deviations and extend the a priori model with quantitative information such as frequencies and performance data. However, a process model yielded by process mining techniques is more similar to a representation of the process behaviour rather than an actual model of the process: it often consists of a huge number of states and interconnections between them, thus resulting in a spaghettilike net which is hard to interpret or even read.

In this paper we propose a novel technique, which we call model mining, to derive an abstract but concise and functionally structured model from event logs. Such a model is not a representation of the unfolded behaviour, but comprises, instead, a set of formal rules for generating the system behaviour. The set of rules is inferred by sifting a plausible a priori model using the event logs as a sieve until a reasonably concise model is achieved (refinement mining). We use rewriting logic as the formal framework in which to perform model mining and implement our framework using the MAUDE rewrite system. Once the final formal model is attained, it can be used, within the same rewriting logic framework, to predict future evolutions of the behaviour through simulation, to carry out further validation or to analyse properties through model checking. We illustrate our approach on a case study from the field of ecology.