Distinguish the Indistinguishable: Spatial Personalized Transformer for Traffic Flow Forecast

In recent years, Spatial-Temporal Graph Neural Networks (STGNNs) has faced increasing challenges in traffic flow forecasting. The main issue lies in the significant indistinguishability among traffic nodes, where highly similar inputs correspond to completely different outputs, posing a major challenge in achieving accurate model fitting. To address this, some researchers have abandoned the graph network structure and introduced spatial embedding to enhance the differentiation between nodes, showing that graph networks are not well-suited for spatial-temporal prediction problems. However, STGNNs are not beyond redemption and the key is indistinguishability. In this paper, we tackle the indistinguishability problem from three main stages of STGNNs: (1) data preprocessing, (2) temporal representation, and (3) graph learning. We discover that the standard attention-based graph learning in (3) and the widely adopted normalization operation in (1) both may worsen data indistinguishability from spatial dimension. Additionally, in temporal dimension, the lack of representation capability also limits the model’s ability to differentiate these indistinguishable data patterns. Base on these findings, we propose the Spatial Personalized Transformer (SPFormer) with optimizations from the three stages mentioned above. Through extensive experiments, our method achieves state-of-the-art performance across multiple publicly available datasets.