Neural Collaborative Filtering to Detect Anomalies in Human Semantic Trajectories

Sumamry:

This study introduces a novel approach to detect anomalies in human semantic trajectories by applying Neural Collaborative Filtering (NCF), which leverages collaborative filtering and neural networks. NCF captures the latent spatial-temporal patterns of human mobility to identify deviations from normal behavior, making it effective even with sparse or incomplete data. This method has applications in surveillance, public health, and safety, such as early infectious disease detection and elder monitoring.

Methodology:

The model uses two core components: a collaborative filtering module to capture routine user visits to points of interest (POIs) and a neural module that analyzes complex spatial-temporal relationships within the data. By estimating a “surprise” score for unexpected movements, the model flags anomalies, and is tested across real-world and simulated datasets. The NCF model operates without semantic data, learning based on visit patterns.

Results:

Testing on simulated and real-world datasets demonstrated that the NCF model outperforms traditional anomaly detection methods, particularly with sparse data, with enhanced detection of spatial anomalies. The results showed high recall rates for detecting anomalies with greater spatial changes but had limited accuracy with smaller datasets due to sparsity.

Conclusion:

The NCF approach shows promise in detecting abnormal behavior patterns in human trajectories. Future work will focus on integrating geographic factors and improving performance in highly sparse datasets by adding synthetic data to bolster robustness.

Full Paper: