This post is by Mohammadreza Fani Sani, Scientific Assistant in the Process And Data Science group at RWTH Aachen University. Contact him via email for further inquiries.
The main aim of process mining is to increase the overall knowledge of business processes. This is mainly achieved by 1) process discovery, i.e. discovering a descriptive model of the underlying process, 2) conformance checking, i.e. checking whether the execution of the process conforms to a reference model and 3) enhancement, i.e. the overall improvement of the view of the process, typically by enhancing a process model. In each case the event data, stored during the execution of the process, is explicitly used to derive the corresponding results.
Many process mining algorithms assume that event data is stored correctly and completely describes the behavior of a process. However, real event data typically contains noisy and infrequent behaviour. The presence of outlier behaviour makes many process mining algorithms, in particular, process discovery algorithms, result in complex, incomprehensible and even inaccurate results. Therefore, to reduce these negative effects, in process mining projects, often a preprocessing step is applied that aims to remove outlier behaviour and keep good behaviour. Such preprocessing phase increases the quality and comprehensiveness of possible future analyses. Usually this step is done manually, which is costly and time-consuming and also needs business/domain knowledge of the data.
In this paper, we focus on improving process discovery results by applying an automated event data filtering, i.e., filtering the event log prior to apply any process discovery algorithm, without significant human interaction. Advantaging from sequential rules and patterns, long distance and indirect flow relation will be considered. As a consequence, the proposed filtering method is able to detect outlier behaviour even in event data with lots of concurrencies, and long-term dependency behaviour. The presence of this type of patterns is shown to be hampering the applicability of previous automated general purpose filtering techniques.
By using the ProM based extension of RapidMiner, i.e., RapidProM, we study the effectiveness of our approach, using synthetic and real event data. The results of our experiments show that our approach adequately identifies and removes outlier behaviour, and, as a consequence increases the overall quality of process discovery results. Additionally, we show that our proposed filtering method detects outlier behaviour better compared to existing event log filtering techniques for event data with heavy parallel and long-term dependency.
