Spatial-Temporal Graph Discriminant AutoEncoder for Traffic Congestion Forecasting

Traffic congestion is a growing issue in modern cities, with significant negative impacts on the environment, the economy, and people's daily lives. Accurately predicting congestion is crucial for effective road control and route planning, making it an essential component of intelligent transportation systems. In this paper, we propose a novel algorithm, the Spatial-Temporal Graph Discriminant Autoencoder (STGDAE), for improving congestion prediction. STGDAE combines graph convolution layers and recurrent neural networks to extract spatial and temporal features from traffic data efficiently. We introduce a distance loss term to improve the autoencoder's feature extraction effectiveness and utilize labels to retain more useful information for congestion prediction. Our extensive experiments on two real-world datasets demonstrate that STGDAE outperforms state-of-the-art methods, achieving an improvement of 0.1 in F1 score on the PeMSD8 dataset. The proposed algorithm has promising potential for improving traffic management in real-world scenarios, such as reducing travel times and fuel consumption and enhancing road safety.