Clustering Large-Scale Origin-Destination Pairs: A Case Study for Public Transit in Beijing

With the extensive collection of various trajectories, a lot of trajectory mining methods have been developed and brought into effect in different applications. The same is true for trajectory clustering. It enables the construction of diverse applications (e.g., mobile social networks) and can promote the intelligence of existing services (e.g., optimizing public transit). In the paper, we propose a three-phase clustering strategy ODTC (Origin Destination pair oriented Trajectories Clustering) for the massive trajectories in the form of OD (Origin Destination) pairs and demonstrate the impact of trajectory clustering on evaluating and adjusting public transit operations. In our ODTC strategy, trajectories are partitioned in the first phase by coarsegrained clustering, reflecting an idea of divide and conquer. While during the second phase of fine-grained clustering, we model the relations of OD pairs as a sparse graph where the spatial and temporal features as well as the constraints of road networks are integrated into the similarity of trajectories. Then we apply a spectral clustering algorithm on the graph to capture clusters. In particular, in the third phase, we borrow the idea from text data mining and give a feasible method to mine the semantics of clusters. As a case study, we perform ODTC on the large-scale trajectories from the Beijing Public Transport Group. From the clustering results, we can observe the mobility patterns of bus passengers. Further, we exploit the clustering results to discover the dynamics of bus operations, evaluate the bus lines and provide support for making the decisions on bus operations.