Transferable Unsupervised Outlier Detection Framework for Human Semantic Trajectories

Sumamry:

This study introduces TOD4Traj, a transferable outlier detection framework designed to identify outliers in human semantic trajectories, which combine spatial, temporal, and textual information. Traditional methods are limited by heuristic rules and lack transferability across datasets. TOD4Traj unifies data features and uses contrastive learning to distinguish both individual and collective movement patterns, enabling unsupervised identification of atypical behaviors in various datasets.

Methodology:

TOD4Traj consists of three main modules:

1. Modality Feature Unification: Aligns spatial-temporal and semantic data, facilitating seamless data integration and increasing the model's adaptability across datasets.
2. Contrastive Learning Module: Identifies regular patterns by recognizing temporal consistency in user behavior, crucial for detecting outliers in individual patterns and across populations.
3. Outlier Quantification: Measures deviations by comparing individual behavior against both historical data and collective trends, determining the degree of outlier status.

Results:

Tests across six datasets demonstrated that TOD4Traj outperforms other methods, excelling in diverse environments. The framework was particularly robust in detecting work and hunger outliers but showed limitations with social outliers. Its transferability across datasets proved effective, reducing the need for retraining on new data.

Conclusion:

TOD4Traj shows potential for real-world applications like public health monitoring and elderly care by identifying unusual trajectories. Future work will address enhanced detection during holiday patterns and refine transfer capabilities for various types of behavioral outliers.

Full Paper: