TGCAE: temporal graph convolutional autoencoder for robust anomaly detection in object-centric process mining

Abstract

With the deep integration of machine learning and cybernetics technologies in business process management (BPM), robust anomaly detection in object-centric process mining is crucial for enhancing operational efficiency and quality within enterprises. Existing anomaly detection approaches predominantly rely on flattened event logs, which overlook multi-object interactions and their complex dependency relationships, thereby limiting the capability to model intricate business processes. In contrast, Object-Centric Event Logs (OCELs) associate events with multiple objects, enabling the modeling of multi-object interactions and complex dependencies inherent in real-world processes. However, current Graph Neural Network (GNN)-based anomaly detection methods primarily focus on structural features while neglecting the critical temporal dynamics of process execution. To bridge this gap, this paper proposes a Temporal Graph Convolutional AutoEncoder (TGCAE) for robust anomaly detection in object-centric process mining. TGCAE effectively captures structural dependencies between process execution instances via Graph Convolutional Networks (GCNs), and integrates a dedicated temporal encoder with an adaptive gating mechanism to model complex temporal patterns. For anomaly detection, a data-driven Interquartile Range (IQR) method is employed to adaptively determine decision thresholds, enabling accurate identification of three typical anomaly types in event logs: attribute manipulation, timestamp perturbation, and activity insertion. Comprehensive experiments on both real and synthetic datasets demonstrate that TGCAE significantly outperforms traditional Autoencoder (AE), Long Short-Term Memory Autoencoder (LSTMAE), and Graph Convolutional Autoencoder (GCNAE) baselines, validating the efficacy of spatiotemporal information fusion for robust anomaly detection.