DSSA-TCN: Exploiting adaptive sparse attention and diffusion graph convolutions in temporal convolutional networks...

Area of Science:

• Intelligent Transportation Systems
• Data Science
• Network Analysis

Background:

• Accurate traffic flow forecasting is crucial for intelligent transportation systems.
• Existing methods often decouple spatial and temporal learning, leading to inefficiencies.
• Challenges include nonlinear dynamics and complex spatio-temporal dependencies in urban networks.

Purpose of the Study:

• To propose a unified framework, DSSA-TCN, for improved spatio-temporal traffic flow forecasting.
• To address limitations of decoupled spatial and temporal learning in existing models.
• To enhance directional interpretability and computational efficiency in traffic prediction.

Main Methods:

• Developed DSSA-TCN, a unified framework with an alternating spatio-temporal coupling mechanism.
• Integrated temporal convolutional blocks with an adaptive spatial module combining sparse attention and diffusion graph convolution.
• Utilized gated dilated convolutions for temporal pattern modeling and bidirectional diffusion convolution for spatial propagation.

Main Results:

• DSSA-TCN demonstrated superior forecasting accuracy and computational efficiency across six real-world datasets.
• The framework provided interpretable spatial reasoning capabilities.
• Achieved significant improvements over existing graph-based and attention-based approaches.

Conclusions:

• Layer-wise coupling of adaptive sparsity and diffusion within a causal temporal backbone is effective.
• DSSA-TCN offers a scalable and physically grounded paradigm for spatio-temporal traffic prediction.
• The proposed method enhances the reliability and interpretability of traffic flow forecasting.