Structure-Aware Graph Multi-Task Learning for Dynamic Sparse OD Demand Prediction

Researchers introduce SAGMTL, a graph-based machine learning framework that improves Origin-Destination demand prediction for transportation systems by jointly modeling regional activity states and flow intensity. The approach addresses real-world challenges of sparse, irregular traffic patterns that existing single-task regression methods struggle to handle, demonstrating superior performance across three major Chinese cities.

SAGMTL tackles a fundamental challenge in intelligent transportation systems: predicting sparse Origin-Destination flows where many potential routes carry zero or minimal traffic. Traditional approaches treating OD prediction as simple regression fail to distinguish between inactive connections and low-demand active ones, limiting accuracy for intermittent routes. The framework's multi-task learning approach—decomposing the problem into structural state modeling and flow intensity estimation—aligns with how transportation networks actually function, where regional activity patterns and connection states drive demand.

This research emerges from growing recognition that sparse, long-tailed data patterns dominate real-world transportation networks, yet most existing methods optimize for average performance rather than low-frequency routes. The node-edge collaborative representation module represents technical sophistication in capturing spatial relationships and temporal dynamics simultaneously, addressing limitations of simpler graph neural networks. Testing across Beijing, Chengdu, and Nanjing datasets ensures findings reflect diverse urban mobility patterns rather than single-city idiosyncrasies.

For transportation technology developers and city planners, this work offers practical improvements in routing optimization, congestion prediction, and resource allocation. More accurate OD demand modeling enables better last-mile delivery routing, improved ride-sharing matching, and reduced urban congestion. The emphasis on robustness in sparse scenarios directly benefits emerging mobility services operating in underutilized corridors.

• →Multi-task learning framework jointly models OD connection activity and flow intensity, improving predictions for sparse traffic patterns.
• →Node-edge collaborative representation captures spatial relationships and temporal dynamics more effectively than single-task regression approaches.
• →Superior performance demonstrated on three major Chinese city datasets suggests framework generalizes across diverse urban environments.
• →Explicit modeling of regional activity states improves robustness when predicting intermittent, low-frequency OD connections.
• →Research enables better urban mobility optimization for routing, congestion management, and transportation resource allocation.