Unsupervised Event Abstraction in a Process Mining Context: A Benchmark Study

Abstract

Due to the rise of IoT, event data becomes increasingly fine-grained. Faced with such data, process discovery often produces incomprehensible spaghetti-models expressed at a granularity level that doesn't match the mental model of a business user. One approach is to use event abstraction patterns to transform the event log towards a more coarse-grained level and to discover process models from this transformed log. Recent literature has produced various (partial) implementations of this approach, but insights how these techniques compare against each other is still limited.

This paper focuses on the use of Local Process Models and Combination based Behavioural Pattern Mining to discover event abstraction patterns in combination with the approach of Mannhardt et al. to transform the event log. Experiments are conducted to gain insights into the performance of these techniques. Results are very limited with a general decrease in fitness and precision and only a minimal improvement of complexity. Results also show that the combination of the process discovery algorithm and the event abstraction pattern miner matters. In particular, the combination of Local Process Models with Split Miner seems to improve precision.