ASTGCL: Adaptive Spatio-Temporal Graph-enhanced Contrastive Learning for Traffic Flow Prediction

With the rise of intelligent urban systems, traffic flow prediction has emerged as a vital element within intelligent transportation frameworks. Graph neural networks (GNNs) are crucial for traffic flow prediction due to the non-Euclidean nature of traffic networks. Most GNN-based approaches depend on predefined adjacency matrices, frequently failing to model complex global dependencies. Additionally, these methods primarily use supervised learning, which is highly dependent on data quality. However, traffic data often suffers from missing values caused by sensor malfunctions, hindering effective feature extraction. To overcome these limitations, a novel framework called Adaptive Spatio-Temporal Graph-enhanced Contrastive Learning for Traffic Flow Prediction (ASTGCL) is introduced to improve prediction performance. Specifically, ASTGCL implements two data augmentation strategies: first, it employs an adaptive, learnable adjacency matrix to enhance the predefined matrix, enabling the model to capture both local and global topological features; second, it integrates three traffic data augmentation techniques to reduce the influence of data noise on prediction accuracy. The enhanced adjacency matrix and augmented traffic data are then utilized in a spatio-temporal contrastive learning process to extract higher-order spatio-temporal features from the traffic flow data. Experiments conducted on four real-world datasets reveal that ASTGCL surpasses baseline models, confirming the efficacy of the proposed framework.