Vehicle Trajectory Clustering via Graph Transformer Networks and Contrastive Learning

Recent advancements in ubiquitous mobile connectivity enable large-scale vehicle trajectory data collection. Such data provide a base for analyzing spatial and temporal dynamics that are essential in transportation, urban planning, and location-based services. Therefore, trajectory representation learning is very helpful for the Internet of Vehicles (IoV). It provides unprecedented insights into movement patterns. This paper proposes Graph Transformer Network (GTN) to perform trajectory embedding. The capabilities of graph neural networks and transformer architectures are exploited to capture both spatial and temporal dependencies in trajectory data. The performance of GTN is assessed in comparison to baseline methods, such as DynWalks, TrajBERT, and DeepWalk, on the T-Drive dataset. Furthermore, clustering is performed by combining GTN with K-Means clustering and GTN with Gaussian Mixture Models (GMM) algorithms. Experimentation on sample data shows that the proposed method outperforms the existing approaches in clustering tasks and captures the complex dynamics in trajectory data.