Multi-perspective hotel operation process anomaly prediction method based on graph transformer and autoencoder

Due to the complexity of hotel operation processes, abnormal situations are inevitable, making proactive anomaly prediction essential for ensuring operational stability. Although current deep learning methods can encode control and data flows to predict anomalies in attributes like activity and time, they often fail to adequately represent the behavioral relationships between activities and lack specific mechanisms to model the interaction between control and data flows. To address these challenges, this paper proposes a business process anomaly prediction method based on a Multi-perspective Graph Transformer and Auto Encoder (MLGTAE). The proposed method first constructs multi-perspective trace graphs by combining Petri nets-which capture process behaviors-with data attributes such as time and resources. It then leverages an attention mechanism to achieve deep semantic interaction between process behavior and data, followed by a decoder that performs reconstruction to detect anomalies. Validated on multiple real-world datasets, the results demonstrate that MLGTAE outperforms existing state-of-the-art methods, showing superior accuracy in predicting anomalies at both the activity and data attribute levels.