Spatial-Temporal Bipartite Graph Attention Network for Traffic Forecasting

Accurate traffic forecasting is pivotal for an efficient data-driven transportation system. The intricate nature of spatial-temporal dependencies and non-linearity present in traffic data has posed a significant challenge to the modeling of accurate traffic forecasting systems. Lately, there has been a significant effort to develop complex Spatial-Temporal Graph Neural Networks (STGNN) that predominantly utilize various Graph Neural Networks (GNN) and attention-based encoder-decoder architectures due to their ability to capture non-linear dependencies in spatial and temporal domains effectively. However, conventional GNNs limit explicit propagation of past information among nodes, while attention-based models such as transformers do not support finer-grained attention score distribution. In this study, we address the aforementioned issues and introduce a novel STGNN namely, Spatio-Temporal Bipartite Graph Attention Network (STBGAT) that allows explicit modeling of past information propagation among nodes. Further, we present a heterogeneous cross-attention mechanism in a transformer to compute finer-grained feature-wise attention distribution enabling the model to capture richer and more expressive temporal dependencies. Our experiments reveal that the proposed architecture outperforms the state-of-the-art approaches proposed in recent literature.