Exploring the Context Generalizability in Spatiotemporal Crowd Flow Prediction: Benchmark and Guideline

Liyue Chen, Xiaoxiang Wang, Leye Wang

Contextual features are important data sources for building spatiotemporal crowd flow prediction (STCFP) models. However, the difficulty of applying context lies in the unknown generalizability of both contextual features (e.g., weather, holiday, and points of interests) and context modeling techniques across different scenarios. In this paper, we build a benchmark composed of large-scale spatiotemporal crowd flow data, contextual data, and state-of-the-art spatiotemporal prediction models. We conduct a comprehensive experimental study to quantitatively investigate the generalizability of different contextual features and modeling techniques in several urban crowd flow prediction scenarios (including bike flow, metro passenger flow, electric vehicle charging demand and so on). In particular, we develop a general taxonomy of context modeling techniques based on extensive investigations in prevailing research. With millions of records and rich context data, we have trained and tested hundreds of different models. Our results reveal several important observations: (1) Using more contextual features may not always result in better prediction with existing context modeling techniques; in particular, the combination of holiday and temporal position can provide more generalizable beneficial information than other contextual feature combinations. (2) In context modeling techniques, using a gated unit to incorporate raw contextual features into the state-of-the-art prediction model has good generalizability. Besides, we also offer several suggestions about incorporating contextual factors for practitioners who want to build STCFP applications. From our findings, we call for future research efforts devoted to developing new context processing and modeling solutions to fully exploit the potential of contextual features for STCFP.

Knowledge Graph



Sign up or login to leave a comment