Process Mining Encoding via Meta-learning for an Enhanced Anomaly Detection

Abstract

Anomalous traces diminish the event log’s quality due to bad execution or security issues, for instance. Focusing on mitigating this phenomenon, organizations spend efforts to detect anomalous traces in their business processes to save resources and improve process execution. Conformance checking techniques are usually employed in these situations. These methods rely on the comparison of the event log obtained and the designed process model. However, in many real-world environments, the log is noisy and the model unavailable, requiring more robust techniques and expert assistance to perform conformance checking. The considerable number of techniques and reduced availability of experts pose an additional challenge to detecting anomalous traces for particular event log scenarios. In this work, we combine the representational power of encoding with a Meta-learning strategy to enhance the detection of anomalous traces in event logs towards fitting the best discriminative capability between common and irregular traces. Our method extracts meta-features from an event log and recommends the most suitable encoding technique to increase the anomaly detection performance. We used three encoding techniques from different families, 80 log descriptors, 168 event logs, and six anomaly types for experiments. Results indicate that event log characteristics influence the representational capability of encodings differently. Our proposed Meta-learning method outperforms the baseline reaching an F-score of 0.73. This performance demonstrates that traditional process mining analysis can be leveraged when matched with intelligent decision support approaches.

The authors would like to thank CNPq (National Council for the Scientific and Technological Development) for their financial support under Grant of Project 420562/2018-4 and 309863/2020-1 and the program “Piano di sostegno alla ricerca 2020” funded by Università degli Studi di Milano.