A network distance and graph-partitioning-based clustering method for improving the accuracy of urban hotspot detection

Abstract

Clustering is an important approach to identifying hotspots with broad applications, ranging from crime area analysis to transport prediction and urban planning. As an on-demand transport service, taxis play an important role in urban systems, and the pick-up and drop-off locations in taxi GPS trajectory data have been widely used to detect urban hotspots for various purposes. In this work, taxi drop-off events are represented as linear features in the context of the road network space. Based on such representation, instead of the most frequently used Euclidian distance, Jaccard distance is calculated to measure the similarity of road segments for cluster analysis, and further, a network distance and graph-partitioning-based clustering method is proposed for improving the accuracy of urban hotspot detection. A case study is conducted using taxi trajectory data collected from over 6500 taxis during one week, and the results indicate that the proposed method can identify urban hotspots more precisely.

Keywords:

• Network space
• graph-partitioning-based clustering
• hotspot detection
• taxi trajectory
• spatiotemporal variations

Acknowledgement

Acknowledges the funding support from several project: an Area of Excellence project (1-ZE24), a startup project (1-ZE6P), and a National Natural Science Foundation of China (No.41622108).