MP-Former: Transformer-Based Online Multi-Task Anomaly Detection

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Detecting anomalies in business processes is essential for maintaining operational efficiency and preventing fraud or misuse. Although process mining enables automated anomaly detection (AD), most methods provide only binary outputs and struggle to distinguish between different anomaly perspectives, handle variable-length event sequences, or detect numerical anomalies—especially in online settings where early detection is critical. This work presents MP-Former, a Transformer-based AD model designed for online, multi-perspective anomaly detection. Using a multi-task learning architecture, MP-Former simultaneously predicts multiple anomaly perspectives with high detail and computational efficiency. We further introduce two novel perspectives—arrival, time and workload, capturing the temporal and inter-case dependencies, which are often overlooked in prior research. Experimental evaluation shows that MP-Former achieves superior F1-scores across all perspectives while maintaining real-time performance, demonstrating the value of specialized multi-perspective approaches for accurate and efficient process anomaly detection.