Trip2Vec: a deep embedding approach for clustering and profiling taxi trip purposes

With the wide availability of GPS trajectory data, sustainable development on understanding travel behaviors has been achieved in recent years. But relatively less attention has been paid to uncovering the trip purposes, i.e., why people make the trips. Unlike to the GPS trajectory data, the trip purposes cannot be easily and directly collected on a large scale, which necessitates the inference of trip purposes automatically. To this end, in this paper, we propose a device-free and novel model called Trip2Vec, which consists of three components. In the first component, it augments the context on trip origins and destinations, respectively, by extracting the information about the nearby point of interest configurations and human activity popularity at particular time periods (i.e., activity period popularity) from two crowdsourced datasets. Such context is well-recognized as the clear clue of trip purposes. In the second component, on the top of the augmented context, a deep embedding approach is developed to get a more semantical and discriminative context representation in the latent space. In the third component, we simply adopt the common clustering algorithm (i.e., K-means) to aggregate trips with similar latent representation, then conduct trip purpose interpretation based on the clustering results, followed by understanding the time-evolving tendency of trip purpose patterns (i.e., profiling) in the city-wide level. Finally, we present extensive experiment results with real-world taxi trajectory and Foursquare check-in data generated in New York City (NYC) to demonstrate the effectiveness of the proposed model, and moreover, the obtained city-wide trip purpose patterns are quite consistent with real situations.