STMAE: Spatial Temporal Masked Auto-Encoder for Traffic Forecasting

The advancement of intelligent transportation systems underscores the importance of data-driven approaches in traffic forecasting, which plays a crucial role in tasks such as traffic signal control and route guidance, among others. However, the inherent uncertainty stemming from regional traffic dynamics, coupled with intricate spatio-temporal correlations, poses formidable challenges to accurate traffic prediction. Moreover, the complexities inherent in sequence forecasting across varying scales further exacerbate the accuracy dilemma. Recognizing the need for integrating information across spatial and temporal dimensions to enhance forecasting precision, a novel solution termed Spatial Temporal Masked Autoencoder (STMAE) is introduced. The STMAE framework addresses these challenges through a two-stage learning process. In the pre-training phase, an autoencoder architecture is employed to extract spatio-temporal features from the data. In the fine-tuning phase, the pre-trained encoder of the STMAE model undergoes further refinement to specifically target traffic forecasting tasks. Extensive evaluations validate the effectiveness of the proposed STMAE model. Notably, STMAE demonstrates competitive performance, achieving 3.32 Vehs MAE for long-term (60 min) traffic forecasting while operating within a reduced computational budget.