MHG-Predictor: A Multi-Layer Heterogeneous Graph-based Predictor for Next Activity in Complex Business Processes

Accurate prediction of the next activity in business processes is crucial for operational efficiency and decision making in modern enterprise systems. Existing methods often fail to capture the multi-layered structural dependencies inherent in event logs. To address this limitation, we propose the Multi-layer Heterogeneous Gaph-based Predictor (MHG-Predictor), where the key innovation lies in our multi-layer heterogeneous graph that decomposes process knowledge into four interconnected layers: global process dynamics, instance-level patterns, activities, and contextual attributes, thereby capturing both temporal and structural dependencies in process execution. To leverage this rich structural information, we develop a multi-layer graph convolutional network (GCN) that combines GatedGraphConv and HeteroSAGE, unified through the multi-layer perceptrons (MLPs) for next activity prediction. Extensive experiments on real-world datasets demonstrate that MHG-Predictor outperforms existing methods, with an average improvement of more than 2% across all evaluation metrics compared to the best baseline method. Furthermore, the four-layer heterogeneous graph has been successfully deployed in the Remote Banking Center's intelligent ticket generation system, achieving an accuracy of up to 80%, representing an approximately 5% improvement over the previous system. The source code is available at https://github.com/Joker3993/MHG-Predictor.