Reinventing Node-centric Traffic Forecasting for Improved Accuracy and Efficiency

Traffic forecasting is a crucial application in smart city efforts. After revisiting the existing literature on deep learning-based traffic forecasting methods, we identify two primary research approaches: node-centric and graph-centric. Node-centric methods focus on constructing spatial features through preprocessing and modeling spatial correlations in the input space. In contrast, graph-centric methods mainly rely on graph neural networks to capture spatial correlations in the latent space. We perform empirical evaluations to identify the pros and cons of each: node methods excel in efficiency while graph methods demonstrate better performance. Based on this, we propose a simple yet effective node-centric framework, named SimST, which overcomes the drawbacks of node-centric methods and enhances their efficiency. Extensive experiments show that SimST achieves performance on par with graph-centric methods while exhibiting up to 39 times inference speedup.