A Novel Transformer-Based Approach for Detecting Anomalies in Event Logs of Process-Aware Information Systems
Eman Abd El-Aziz, Radwa Fathalla, Yasser Ismail, Mohamed Shaheen · 2023
Anomaly detection in the logs generated from the Process-Aware Information Systems is a critical task to prevent failures that can harm the performance of any organization. The majority of the anomaly detection approaches applied in the business process domain are unsupervised approaches because of the lack of labeled anomaly data and Perform poorly since they have no prior knowledge of the anomalies. Furthermore, almost all of them utilize a threshold to distinguish normal process executions from anomalous ones. Therefore, they have to change the threshold each time they use it on a different test set and need to struggle to find the optimal threshold that maximizes the accuracy for all test sets. This is so inappropriate to apply in real situations. In this study, we present a new method that is based on transformers to address the low-performance limitation of the current unsupervised approaches. Additionally, we propose to utilize the logistic regression classifier to separate anomalies from the data outputted based on the transformer to avoid using a threshold and enhance detection accuracy. To assess the proposed method, we used two real life and two synthetic event logs. The experimental results showed that the proposed method outperforms six competing anomaly detection methods.