Author:

Abstract:

Process mining is the research area that is concerned with knowledge discovery from event logs, which are recorded and stored by a wealth of information systems, used to support, model and govern operational processes. Event logs can contain up to millions of events, and organizations face the challenge of extracting value from this vast data repository, so as to improve their business processes by learning from insights derived from these data sets. Typically, the field of process mining is structured by categorizing its analysis tasks into a taxonomy containing three broad task types: process discovery, conformance checking, and process enhancement. Although great strides have been made in the field of process mining towards improving business processes and deriving insights from historical event-based data repositories, the field faces some notable difficulties. One particular difficulty is that process mining, with process discovery in particular, is frequently limited to the setting of unsupervised learning, as negative information (i.e. events that were prevented from taking place) are often unavailable in real-life event logs. Multiple solutions have been proposed to resolve this aspect, with one of them encompassing the artificial induction of negative events based on available, positive, information. This concept is put forward as the topic of focus throughout the first part of this thesis. We outline an improved artificial negative event generator, and subsequently enhance this technique with a scoring mechanism to assign a measure of confidence to induced negative events. Next, we show how the induced negative events can be applied in a conformance checking setting, as they allow us to develop a comprehensive conformance checking framework in line with standard machine learning practices, allowing to assess the recall, precision and generalization of a process model. Finally, we also show how artificial negative events can be applied to uncover implementation problems by using them as highlighters of unobserved behavior. The second part of this thesis outlines a number of additional contributions which are not directly related to the concept of artificial negative events. In particular, a novel heuristic process discovery technique is presented, based on a long lineage of related process discovery algorithms, but with a particular focus on robustness and flexibility. Next, based on replay strategies developed throughout the first part of the thesis, an event-granular conformance analysis technique is presented which can be applied towards enabling real-time monitoring of business activities. Next, a technique for explaining event log cluster solutions on an instance-granular level is presented. Finally, a benchmarking framework is developed to enable the automated set-up of large-scale conformance analysis experiments.