Multi-Graph Anomaly Detection in Business Processes With Scalable Neural Architectures

Abstract

Business Process Management (BPM) solutions are critical for organizational efficiency, but their potential remains limited by inadequate effectiveness in anomaly detection capabilities for real-world deployment. This study addresses key challenges in developing production-ready anomaly detection systems that are scalable, efficient, and adaptable across diverse business domains. We propose several enhancements to a state-of-the-art graph-based autoencoder model to overcome these barriers. This includes improved artificial anomaly injection methods that more accurately reflect real-world scenarios to overcome the scarcity of annotated datasets in real-world environments. A comprehensive study of multiple model architectures is conducted, incorporating Graph Attention v2 in the encoder and replacing Gated Recurrent Unit (GRU) decoders with Transformers, thereby achieving comparable or superior performance with half the computational cost. Introducing a denoising objective alongside reconstruction, we lay the foundation for targeted training on domain-specific anomalies without compromising general detection capabilities. We demonstrate the solution's reliability and generalizability in varied business domains by conducting comprehensive evaluations on diverse public and private datasets. The results indicate significant improvements in scalability and real-world applicability while maintaining and enhancing detection accuracy, with results showing up to 22% increase in anomaly detection performance.