[Paper Review] Short-term passenger flow prediction for multi-traffic modes: A Transformer and residual network based multi-task learning method

Jian Yang, Jinlei Zhang+3 more|arXiv (Cornell University)|Feb 27, 2022

Traffic Prediction and Management Techniques4 citations

TL;DR

This paper proposes Res-Transformer, a multi-task learning model combining modified Transformer layers and residual networks to predict short-term passenger flows across subway, taxi, and bus modes. By jointly modeling spatiotemporal dependencies and cross-mode correlations, it achieves superior accuracy and robustness on real-world Beijing datasets compared to baselines.

ABSTRACT

With the prevailing of mobility as a service (MaaS), it becomes increasingly important to manage multi-traffic modes simultaneously and cooperatively. As an important component of MaaS, short-term passenger flow prediction for multi-traffic modes has thus been brought into focus. It is a challenging problem because the spatiotemporal features of multi-traffic modes are critically complex. Moreover, the passenger flows of multi-traffic modes differentiate and fluctuate significantly. To solve these problems, this paper proposes a multitask learning-based model, called Res-Transformer, for short-term inflow prediction of multi-traffic modes (subway, taxi, and bus). Each traffic mode is treated as a single task in the model. The Res-Transformer consists of two parts: (1) several modified Transformer layers comprising the conv-Transformer layer and the multi-head attention mechanism, which helps to extract the spatial and temporal features of multi-traffic modes, (2) the structure of residual network is utilized to obtain the correlations of different traffic modes and prevent gradient vanishing, gradient explosion, and overfitting. The Res-Transformer model is evaluated on two large-scale real-world datasets from Beijing, China. One is the region of a traffic hub and the other is the region of a residential area. Experiments are conducted to compare the performance of the proposed model with several baseline models. Results prove the effectiveness and robustness of the proposed method. This paper can give critical insights into the short-term inflow prediction for multi-traffic modes.

Motivation & Objective

• To address the challenge of predicting short-term passenger flows across multiple transportation modes (subway, taxi, bus) in a unified framework.
• To model complex spatiotemporal patterns and inter-mode correlations in multi-traffic systems.
• To improve prediction accuracy and model stability by mitigating gradient vanishing and overfitting in deep learning architectures.
• To evaluate the model on real-world datasets from diverse urban regions—traffic hubs and residential areas.

Proposed method

• The model employs modified Transformer layers with conv-Transformer and multi-head attention mechanisms to extract spatial and temporal features from multi-traffic flow data.
• A residual network structure is integrated to enhance feature learning, stabilize training, and reduce overfitting across tasks.
• Each transportation mode (subway, taxi, bus) is treated as a separate but jointly optimized task in a multi-task learning framework.
• The architecture enables cross-task knowledge transfer by sharing lower-level features while maintaining task-specific heads for individual predictions.
• The model is trained end-to-end on two large-scale real-world datasets from Beijing, covering different urban zones.
• The use of residual connections helps prevent gradient vanishing and explosion during deep network optimization.

Experimental results

Research questions

• RQ1 Can a unified deep learning model effectively predict short-term passenger flows across multiple transportation modes simultaneously?
• RQ2 How do spatiotemporal dependencies and inter-mode correlations impact prediction performance in multi-traffic systems?
• RQ3 To what extent does multi-task learning with shared and task-specific components improve prediction accuracy and robustness?
• RQ4 How does the integration of residual networks enhance training stability and generalization in multi-traffic flow prediction?

Key findings

• The Res-Transformer model outperforms multiple baseline models in both root mean square error (RMSE) and mean absolute error (MAE) across all traffic modes.
• The model demonstrates robust performance on two distinct urban regions—traffic hubs and residential areas—indicating strong generalization ability.
• The multi-task learning setting improves prediction accuracy compared to single-task baselines, especially for low-flow modes like taxis.
• The ablation study confirms that both the modified Transformer and residual network components contribute significantly to performance gains.
• The model effectively captures long-range temporal dependencies and spatial correlations across different transport networks.
• The integration of residual connections reduces training instability and prevents overfitting, particularly in deeper architectures.