CLGSDN: Contrastive-Learning-Based Graph Structure Denoising Network for Traffic Prediction

The graph neural network-based prediction models have demonstrated remarkable utility in traffic prediction, and their efficacy is highly determined by the quality of the provided graphs. Consequently, there is an increasing demand for employing graph structure learning (GSL) techniques to optimize or generate the graphs. However, existing GSL techniques for traffic prediction encounter various issues, including the absence of temporal dynamicity, noisy connections, and insufficient supervisory information. To address these limitations, this article proposes a novel two-stage graph generation framework called contrastive learning-based graph structure denoising network (CLGSDN). This framework formulates the graph generation task as a probabilistic observation-inference process: using the self-learning adjacency matrix and time delayed self-attention (TDSA) methods to generate a series of graph observations, then inferring the optimal graph based on observations. The self-learning adjacency matrix is responsible for learning all potential connections in the graph, while TDSA enables the graph to change with traffic flow. In addition, CLGSDN identifies and eliminates noisy connections by modeling negative samples of the graph (edges), and defines virtual labels to achieve spatiotemporal graph contrastive learning (ST-GCL) in traffic prediction. The experimental results show that CLGSDN significantly enhances current mainstream traffic prediction models by providing reliable and efficient graphs. As such, it has significant implications for a wide range of applications, including traffic management, logistics, and smart transportation systems.