Exploring Context Generalizability in Citywide Crowd Mobility Prediction: An Analytic Framework and Benchmark

TL;DR

This work tackles the generalizability of contextual information in citywide crowd mobility prediction by proposing a unified analytic framework and a large-scale benchmark across bike-sharing, metro, and EV charging tasks. It systematically analyzes both contextual features (weather, holidays, TP, POIs, roads, demographics, SP) and context modeling techniques (Early Joint vs Late Fusion, with 12 late-fusion variants and gating-based methods). Key findings show that more contextual features do not guarantee better performance; combining holiday with temporal position (Holi-TP) and using Raw-Gating provides strong, cross-scenario generalization, while embedding is not always necessary for gating. The study offers actionable guidelines and a lightweight feature-selection strategy, and demonstrates generalizability across new applications and ST models, highlighting the need for novel context processing solutions in spatiotemporal prediction. These insights have practical implications for deploying robust, context-aware city-scale mobility prediction systems and set the stage for further methodological advances in contextual modeling.

Abstract

Contextual features are important data sources for building citywide crowd mobility prediction models. However, the difficulty of applying context lies in the unknown generalizability of contextual features (e.g., weather, holiday, and points of interests) and context modeling techniques across different scenarios. In this paper, we present a unified analytic framework and a large-scale benchmark for evaluating context generalizability. The benchmark includes crowd mobility data, contextual data, and advanced prediction models. We conduct comprehensive experiments in several crowd mobility prediction tasks such as bike flow, metro passenger flow, and electric vehicle charging demand. 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 deep prediction model has good generalizability. Besides, we offer several suggestions about incorporating contextual factors for building crowd mobility prediction applications. From our findings, we call for future research efforts devoted to developing new context modeling solutions.

Figures (16)

• Figure 1: Examples of temporal and spatial contextual features.
• Figure 2: Statistics on recent crowd mobility prediction papers, covering the period up to November 2024.
• Figure 3: Impact of different weather conditions on citywide crowd flow. The red solid line represents a specific weather condition (e.g., snow), while the blue dashed line represents the other conditions in comparison (e.g., non-snow). The greater the difference between the red and blue curves, the more significant the impact of the weather conditions on crowd mobility. Best viewed in color.
• Figure 4: Citywide hourly crowd flow during holidays and workdays.
• Figure 5: Citywide daily crowd flow during a week.