[Paper Review] Multimodal joint prediction of traffic spatial-temporal data with graph sparse attention mechanism and bidirectional temporal convolutional network

Dongran Zhang, Jiangnan Yan+3 more|arXiv (Cornell University)|Dec 24, 2024

Traffic Prediction and Management Techniques16 citations

TL;DR

GSABT introduces a graph sparse attention mechanism with a bidirectional temporal convolutional network to perform multimodal traffic spatial-temporal joint prediction, achieving state-of-the-art results on three real datasets. It jointly models multiple traffic modes and scales to spatial-temporal extensions.

ABSTRACT

Traffic flow prediction plays a crucial role in the management and operation of urban transportation systems. While extensive research has been conducted on predictions for individual transportation modes, there is relatively limited research on joint prediction across different transportation modes. Furthermore, existing multimodal traffic joint modeling methods often lack flexibility in spatial-temporal feature extraction. To address these issues, we propose a method called Graph Sparse Attention Mechanism with Bidirectional Temporal Convolutional Network (GSABT) for multimodal traffic spatial-temporal joint prediction. First, we use a multimodal graph multiplied by self-attention weights to capture spatial local features, and then employ the Top-U sparse attention mechanism to obtain spatial global features. Second, we utilize a bidirectional temporal convolutional network to enhance the temporal feature correlation between the output and input data, and extract inter-modal and intra-modal temporal features through the share-unique module. Finally, we have designed a multimodal joint prediction framework that can be flexibly extended to both spatial and temporal dimensions. Extensive experiments conducted on three real datasets indicate that the proposed model consistently achieves state-of-the-art predictive performance.

Motivation & Objective

• Motivate multimodal traffic prediction across different transportation modes.
• Develop a flexible framework that jointly extracts spatial-temporal features across modes.
• Address sparsity and heterogeneity in multimodal traffic graphs.
• Propose a scalable architecture that extends to both spatial and temporal dimensions.

Proposed method

• Use a multimodal joint graph with self-attention to capture local spatial features.
• Apply Top-U sparse attention for global spatial feature interaction across modes.
• Introduce a Shared Bidirectional Temporal Convolutional Network (S-BiTCN) for common temporal features.
• Develop Unique Bidirectional Temporal Convolutional Networks (U-BiTCN) for modality-specific temporal features.
• Fuse spatial and temporal features through an MLP predictor with residual connections.

Experimental results

Research questions

• RQ1 Can a graph sparse attention mechanism effectively capture both local and global spatial dependencies across multiple traffic modes?
• RQ2 Does a bidirectional temporal convolutional network improve temporal feature learning for multimodal traffic data?
• RQ3 How well does the proposed GSABT framework scale when extending to more modalities and extended spatial-temporal dimensions?
• RQ4 What are the relative gains of shared vs. unique temporal modules for inter- and intra-modal dynamics?

Key findings

• GSABT achieves superior predictive performance across all three real-world datasets (BJ Taxi, NYC Taxi, NYC Bike) compared with baselines.
• On dual-data joint predictions, GSABT reduces MAE and RMSE and increases Pearson correlation (PCC) relative to competing models on multiple tasks.
• Augmented coupling experiments show GSABT benefits from joint modeling across three datasets, with notable MAE/RMSE reductions and PCC gains.
• Decoupling experiments confirm GSABT maintains strong performance when modalities are modeled independently.
• The model demonstrates robust improvements with 2-modal and 3-modal joint predictions, validating its scalability and flexibility.