DCSD‐Net: Density‐Classification, Contrastive and Self‐Distillation Network for Traffic Congestion Prediction

With rapid urbanisation, traffic congestion has become increasingly severe, making accurate congestion prediction essential for improving the operational efficiency of intelligent transportation systems. In recent years, spatio‐temporal graph neural networks have been employed by many researchers to capture spatial and temporal dependencies in traffic data, achieving promising prediction performance. However, these models often have limited feature representation and poor generalisability. To address these issues, this paper proposes a traffic congestion prediction model based on traffic density classification that incorporates contrastive learning and self‐distillation. In the data preprocessing stage, traffic density is computed based on key traffic indicators, with congestion levels categorised accordingly. A temporal feature generator combining a Variational Autoencoder (VAE) with a contrastive learning module is designed to enhance feature diversity. In addition, a self‐distillation mechanism with teacher‐student architecture improves prediction accuracy by transferring informative soft targets. Experimental results on real‐world datasets demonstrate that the proposed method outperforms existing approaches across multiple performance metrics and provides an effective and generalisable solution for traffic congestion prediction.