Mining multi-variant process models from low-level logs

Abstract

Process discovery techniques are a precious tool for analyzing the real behavior of a business process. However, their direct application to lowly structured logs may yield unreadable and inaccurate models. Current solutions rely on event abstraction or trace clustering, and assume that log events refer to well-defined (possibly low-level) process tasks. This reduces their suitability for logs of real BPM systems (e.g. issue management) where each event just stores several data fields, none of which fully captures the semantics of performed activities. We here propose an automated method for discovering an expressive kind of process model, consisting of three parts: (i) a logical event clustering model, for abstracting low-level events into classes; (ii) a logical trace clustering model, for discriminating among process variants; and (iii) a set of workflow schemas, each describing one variant in terms of the discovered event clusters. Experiments on a real-life data confirmed the capability of the approach to discover readable high-quality process models.