A cross-modal transferable approach for comprehensive travel demand prediction

Accurate prediction of comprehensive travel demand-that is, the full travel demand across all population groups and transportation modes-is essential for intelligent transportation systems (ITS). Recently, many studies have explored this task from time series analysis to spatiotemporal modeling, most of which rely on large volumes of high-quality historical data. However, such comprehensive travel demand data are often difficult to obtain in practice, while partial multimodal transportation data are more readily available. In this paper, we consider a more practical scenario where the target region lacks historical data during training, but related source regions with varying levels of data are accessible. To this end, we propose a cross-modal transferable prediction approach that leverages both partial multimodal transportation data and comprehensive travel demand data from source regions to improve performance in target regions lacking sufficient comprehensive travel data. Specifically, our method follows a pretraining–fine-tuning paradigm: during pretraining, the model learns predictive knowledge to facilitate parameter adaptation in fine-tuning, while in fine-tuning, a temporal importance parameter ߙ is introduced to capture domain-invariant knowledge that enhance transferability. Experimental results on real-world datasets show that our method consistently outperforms baselines by approximately 6%, demonstrating its effectiveness in unstable task environments.