NSP-AE: Neuro-symbolic process autoencoder for time anomaly detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Business Process Management

Background:

• Temporal anomaly detection is crucial for identifying deviations in business process execution times.
• Existing methods often overlook joint analysis of process structure and temporal data.
• Abnormal temporal behavior can occur even in control-flow compliant processes.

Purpose of the Study:

• To propose NSP-AE, a neural-symbolic process autoencoder for enhanced temporal anomaly detection.
• To jointly model process structural constraints and temporal deviation signals for accurate anomaly scoring.
• To improve the identification of process instances with abnormal temporal behavior.

Main Methods:

• Representing event traces with activity, resource, and local delta-time information.
• Discovering Petri nets and using prefix replay to determine reachable activities.
• Developing a dynamic compliance mask embedded in the decoder to constrain activity prediction.
• Combining activity prediction deviation, compliance signals, and time-interval deviation for anomaly scoring.

Main Results:

• NSP-AE integrates process structure and temporal data for anomaly scoring.
• The method effectively constrains activity prediction using dynamic compliance masks.
• Experiments demonstrate NSP-AE's superior performance over baseline methods on public event logs.
• Statistically significant advantages were observed on multiple datasets.

Conclusions:

• NSP-AE offers a robust approach to temporal anomaly detection in business processes.
• The neural-symbolic integration effectively leverages both structural and temporal information.
• The proposed method enhances the ability to detect subtle temporal deviations missed by other techniques.