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

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Mining

Background:

• Hotel operations involve complex processes prone to inevitable abnormal situations.
• Current deep learning methods struggle to fully represent activity relationships and control-data flow interactions for anomaly prediction.
• Proactive anomaly prediction is crucial for maintaining operational stability in complex systems.

Purpose of the Study:

• To propose an advanced business process anomaly prediction method.
• To enhance the representation of behavioral relationships and data attribute interactions in anomaly detection.
• To improve the accuracy of predicting anomalies at both activity and data attribute levels.

Main Methods:

• Constructing multi-perspective trace graphs using Petri nets and data attributes (time, resources).
• Employing an attention mechanism for deep semantic interaction between process behavior and data.
• Utilizing an autoencoder for reconstruction-based anomaly detection.

Main Results:

• The proposed Multi-perspective Graph Transformer and Auto Encoder (MLGTAE) method was validated on real-world datasets.
• MLGTAE demonstrated superior performance compared to existing state-of-the-art anomaly prediction methods.
• The method achieved higher accuracy in detecting anomalies at both activity and data attribute levels.

Conclusions:

• MLGTAE effectively addresses limitations in current deep learning approaches for business process anomaly prediction.
• The multi-perspective graph representation and attention mechanism enhance the understanding of complex process dynamics.
• The findings highlight the potential of MLGTAE for improving operational stability through accurate anomaly detection.