Learning Urban Dynamics with Transformers: Joint Prediction of Traffic and Ride Demand

Abstract

Traffic speed and ride-hailing demand are closely intertwined in dense urban areas, yet they are almost always predicted by separate, unconnected models. This paper presents STT-CTA (SpatialTemporal Transformer with Cross-Task Attention), a single architecture that jointly predicts hourly traffic speed and ride-hailing demand on a regular urban grid using factorized spatial-temporal self-attention and a learnable-gated cross-task attention module that enables selective information exchange between tasks. We evaluate the model on one full year of data from Chicago’s Loop district, combining traffic sensor readings, trip records, weather observations, and public holidays into a shared input representation. The key result is an asymmetric directional coupling: letting the demand branch attend to speed signals reduces demand MAPE by 2.2 percentage points and MAE by up to 8.4% at a 6-hour horizon, whereas feeding demand information back into the speed branch degrades speed accuracy. This asymmetry supports a hybrid deployment strategy and carries practical implications for the design of multi-task urban transportation forecasting systems, where different tasks may benefit unequally from shared information.