Hunting image: Taxi search strategy recognition using Sparse Subspace Clustering

This study proposes hunting image, an image-based representation, to describe and recognize taxi passenger-search strategies. Four features are selected to generate an image for each taxi, namely, the cruising speed, the cruising ratio, the difference in demand-supply ratios and the next passenger trip distance. These features can be easily computed for each taxi search trip using widely available data sources. Sparse subspace clustering (SSC), an unsupervised learning algorithm, is introduced to identify search strategies embedded in the hunting images. The proposed methodology is experimented on a large-scale taxi trajectory dataset collected in Shenzhen, China across five months in 2016. Twenty four clusters corresponding to different search strategies are identified from 885 taxis. The differences in search strategies are linked to the operational efficiency and profitability of individual taxis. The results also reveal common search patterns in the taxi market of Shenzhen. Specifically, we find most taxis: (1) prefer to cruise in the same region after dropping off the last passenger; (2) have trouble finding passenger at midnight and early in the morning; and (3) tend to gather at the ports between Shenzhen and Hong Kong during peak periods to serve short-distance trips.