CALM: cluster-aware linear model for dynamic traffic flow prediction

This is the authors' abstract. We don't add key points for this paper.

Traffic flow prediction is a fundamental task in Intelligent Transportation Systems (ITS), essential for alleviating urban congestion, optimizing traffic operations, and enabling autonomous driving. Recent advances in spatio-temporal Transformers and deep learning have led to notable progress in traffic modeling. However, the inherent nonlinearity, non-stationarity, and high dimensionality of traffic data continue to challenge existing models, which often lack robustness and are sensitive to anomalies, limiting their effectiveness in complex real-world scenarios.To address these issues, we propose a novel framework called Cluster-Aware Linear Model (CALM). CALM first integrates the DBSCAN clustering algorithm with a sliding time window to construct dynamic spatio-temporal graphs, enabling adaptive modeling of evolving traffic patterns. It then leverages both local and global correlation matrices to generate rich feature representations that jointly capture short-term fluctuations and long-term trends, effectively mitigating heterogeneity across different time scales. Importantly, to reduce model complexity and inference cost associated with Transformer architectures, CALM replaces the self-attention mechanism with a lightweight linear layer in the prediction stage. This not only enhances computational efficiency but also improves model robustness, making it more suitable for deployment in real-time or resource-constrained environments. Extensive experiments demonstrate that CALM outperforms mainstream prediction models, particularly in handling non-stationary dynamics and sudden traffic events. The source code is publicly available at: https://github.com/chenjing0616/CALM .