Multi-Mode Spatiotemporal Adaptive Fusion Network for Travel Demand Prediction

The development of multi-mode transportation systems, e.g., bus, metro, taxi, and bike-sharing, presents a fundamental challenge in forecasting demand across heterogeneous, noisy, and complexly interacting data streams. From a feature modeling perspective, this requires a shift from simple data fusion to a more principled approach. This paper introduces a novel end-to-end framework, Multi-mode Spatiotemporal Adaptive Fusion Network (MSTAFN), that systematically addresses this challenge through a two-stage process: 1) unsupervised shared feature selection, and 2) dynamic asymmetric feature interaction modeling. For the first stage, we design an Infomax module that employs an information-theoretic principle to obtain a clean low-dimensional shared latent representation from cross-mode data. This representation captures the underlying semantic drivers of demand, such as latent commuting patterns, while mitigating noise and redundancy. For the second stage, we propose a Multi-Flashback module to explicitly model the complex asymmetric interactions between heterogeneous features, particularly across different temporal granularities. Its Cross-Flashback mechanism is designed to allow low-frequency modes (e.g., metro) to be informed by the latest fine-grained dynamics of high-frequency modes (e.g., bike-sharing). Experiments on a large-scale real-world dataset from New York City demonstrate that our two-stage paradigm outperforms state-of-the-art baselines, especially on the mode with the coarsest time granularity. This validates the superiority of our proposed modeling framework, supporting the improvement of operational efficiency for multi-mode transportation systems.