Temporal Hierarchical Graph Attention Network for Traffic Prediction with Prompt Learning

Traffic prediction is a complex task. In the traditional forecasting methodologies, there are limitations regarding generalization, as specifically observed in the pretraining approaches. This is particularly evident in models involving temporal and spatial attributes, where finetuning often encounters numerous challenges. With the emergence of graph prompt learning, a new finetuning paradigm has gained increasing application. In the context of the traffic prediction methods involving spatio-temporal attributes, we endeavor to enhance its performance in the downstream tasks through graph pretraining and prompt-based finetuning. In particular, we propose a traffic prediction model based on graph prompt learning. This model deeply mines the hierarchical information of traffic data and its temporal attributes, by employing the method of graph pretraining and prompt learning to enhance the model’s generalizability and accuracy. Extensive experiments conducted on two real-world traffic prediction datasets confirm the effectiveness of the proposed method.