Process Discovery Enhancement with Trace Clustering and Profiling

Abstract

The potential in process mining is progressively growing due to the increasing amount of event-data. Process mining strategies use event-logs to automatically classify process models, recommend improvements, predict processing times, check conformance, and recognize anomalies/deviations and bottlenecks. However, proper handling of event-logs while evaluating and using them as input is crucial to any process mining technique. When process mining techniques are applied to flexible systems with a large number of decisions to take at runtime, the outcome is often unstructured or semi-structured process models that are hard to comprehend. Existing approaches are good at discovering and visualizing structured processes but often struggle with less structured ones. Surprisingly, process mining is most useful in domains where flexibility is desired. A good illustration is the "patient treatment" process in a hospital, where the ability to deviate from dealing with changing conditions is crucial. It is useful to have insights into actual operations. However, there is a significant amount of diversity, which contributes to complicated, difficult-to-understand models. Trace clustering is a method for decreasing the complexity of process models in this context while also increasing their comprehensibility and accuracy. This paper discusses process mining, event-logs, and presenting a clustering approach to pre-process event-logs, i.e., a homogeneous subset of the event-log is created. A process model is generated for each subset. These homogeneous subsets are then evaluated independently from each other, which significantly improving the quality of mining results in flexible environments. The presented approach improves the fitness and precision of a discovered model while reducing its complexity, resulting in well-structured and easily understandable process discovery results.

A Trace Clustering Framework for Improving the Behavioral and Structural Quality of Process Models in Process Mining

“…In this section, we present the results obtained by application of the proposed framework for trace clustering. We used two baseline trace clustering techniques Jablonski et al [20] and Faizan et al [24] for the experimental evaluation of our approach and refer to them as baseline 1 and baseline 2 throughout this section.…”

“…In the context of clustering the traces based on trace features, researchers in [50], [51], [20], [24], and [52] proposed to use several distinct event log features to group similar traces. The main idea behind trace-features-based clustering was to project the event log into a vector space representation called trace profiles.…”

“…In context of feature-based trace clustering, some researchers [50], [51], [20], [24], [52] we successful in enhancing the fitness quality of the process models. However, equally vital process perspective, the model complexity was disregarded.…”

“…Keeping in view the identified limitations, we present an enhanced trace clustering framework that not only produces more precise but also yields simpler process models. Furthermore, the proposed framework exhibits a notable time reduction of 31 minutes in comparison to the two baseline studies [20] and [24], which required 201 minutes and 306 minutes respectively. Consequently, the proposed framework not only enhances the overall performance but also significantly improves the time efficiency of the trace clustering solution.…”