GADEF-Net: A heterogeneity-aware dual-graph framework for robust multimodal traffic forecasting.

Area of Science:

• Intelligent Transportation Systems (ITS)
• Traffic Forecasting
• Machine Learning

Background:

• Effective traffic forecasting in ITS relies on integrating heterogeneous sensing data.
• Existing graph-based models often use uniform encoders, ignoring varied spatiotemporal propagation patterns of traffic states and contextual signals.

Purpose of the Study:

• To propose a heterogeneity-aware multimodal forecasting framework that addresses limitations in current traffic prediction models.
• To improve the accuracy and adaptability of traffic forecasting by accounting for distinct data stream characteristics.

Main Methods:

• Developed the Adaptive Gated Dual-Graph Network (GADEF-Net), a novel framework for multimodal traffic forecasting.
• Employed a dual-branch architecture: an attention-based branch for global temporal evolution and a diffusion-convolution branch for localized context propagation.
• Integrated representations using an Adaptive Gated Fusion (AGF) module for dynamic weight adjustment.

Main Results:

• GADEF-Net demonstrated strong overall performance across three real-world datasets.
• Achieved leading results on the Daegu-Urban and PeMS08 datasets.
• Maintained competitive performance on the PeMS-BAY dataset, indicating robustness.

Conclusions:

• Explicitly modeling heterogeneous propagation mechanisms enhances multimodal traffic forecasting.
• The proposed approach is particularly beneficial for stochastic traffic conditions and longer prediction horizons.
• While effective, the model introduces additional computational costs for practical implementation.