Mining Mobility Patterns with Trip-Based Traffic Analysis Zones: A Deep Feature Embedding Approach

Intelligent transportation system applications often require mining salient patterns in a massive amount of mobility data. Previous mobility pattern mining methods are mainly based on established Traffic Analysis Zones (TAZ). However, these TAZs only convey spatial geographic properties but are not optimized with trip data for capturing inter-zone mobility patterns. To obtain more precise and semantic mobility patterns, we propose a mobility pattern mining framework that simultaneously computes trip-based TAZs, with the option to incorporate a wide range of auxiliary information. Specifically, we embed spatial geographic properties, taxi trip data and auxiliary attributes into high-dimensional features via deep neural networks. By maximizing soft-HGR, a statistical measure of correlation, most correlated features of trip origins and destinations can be extracted, which are further clustered to obtain trip-based TAZs and inter-zone mobility patterns. In the experiments, we justified the effectiveness of our soft-HGR trip embedding approach by destination prediction on real taxi trip data, improving prediction accuracy by 4.2%. Comparing with other popular methods, our approach achieves the best result in terms of tight TAZs and high origin-destination correlations between zones. We further presented two case studies that incorporate different auxiliary information, point-of-interest attributes and temporal factors, which showed insightful findings for urban transportation management.