Large Language Models for Spatial Trajectory Patterns Mining

Sumamry:

This paper explores the potential of Large Language Models (LLMs) like GPT-4 and Claude-2 to detect anomalous patterns in spatial trajectories, specifically within human mobility data, as a novel approach compared to traditional trajectory anomaly detection methods. By evaluating LLMs’ ability to identify deviations from typical mobility patterns, such as unexpected changes in trajectories, the research sheds light on LLMs’ suitability for diverse applications, from elder care monitoring to disease tracking.

Methodology:

The researchers conducted empirical tests using LLMs, comparing their performance to conventional trajectory anomaly detection models on two datasets: GEOLIFE and PATTERNS-OF-LIFE. The study analyzed LLMs’ performance in different prompt settings, both with and without contextual hints (markers indicating potential anomalies), and tested whether presenting data as a combined input or as separate inputs affected outcomes. Results were benchmarked against several deep learning and traditional methods, providing a comprehensive view of LLMs' anomaly detection capabilities.

Results:

The findings show that LLMs, especially Claude-2 with contextual hints, performed comparably to some state-of-the-art deep learning methods in identifying anomalous trajectories on human mobility data. The study demonstrated that LLMs not only detect unusual trajectory patterns but also offer coherent, interpretable explanations for their anomaly scores. Hint-based prompts further enhanced the accuracy of anomaly detection, particularly for LLMs with longer context windows.

Conclusion:

LLMs demonstrate substantial promise in trajectory anomaly detection, even in cases lacking explicit anomaly markers. The study suggests that LLMs could be integrated into real-world applications requiring interpretability and nuanced detection of behavioral changes, though improvements in context window limitations are needed for analyzing lengthy data. Future work will focus on open-source LLM models, additional datasets, and refining LLM processing capabilities for extended mobility patterns.

Full Paper: