GSASN: a graph self-learning attention scores network for spatial modeling of network traffic matrix prediction

Network traffic prediction is a critical technology for next-generation intelligent routers, enabling network managers to effectively plan resources and address bandwidth and latency challenges posed by rapidly growing data applications and video traffic. While time series models are commonly employed for traffic matrix (TM) prediction, existing spatial modeling approaches-including Convolutional Neural Networks (CNNs), Graph Convolutional Networks (GCNs), and Graph Attention Networks (GATs)-exhibit significant limitations. These models often fail to accurately capture complex nonlinear spatial relationships between network nodes, primarily because current attention score calculation methods rely on input-dependent feature similarities that impose linear constraints in log-space. To address these fundamental limitations, we propose the Graph Self-learning Attention Scores Network (GSASN), which autonomously learns attention scores between same-origin or same-destination nodes through a learnable parameter matrix, effectively breaking the linear constraints of conventional approaches. Experimental results on two real-world datasets demonstrate that GSASN achieves substantial improvements over GAT: up to 24.7% in Root Mean Square Error (RMSE), 36.5% in Mean Absolute Error (MAE), and 7.3% in coefficient of determination (R²). Furthermore, we present a novel spatio-temporal model (ST-GSASN) that integrates GSASN with temporal learning components through gated fusion. Extensive experiments reveal that ST-GSASN outperforms both baseline and state-of-the-art methods, including recent transformer-based approaches (ASTGNN, STTN), while maintaining superior computational efficiency with only 5% of the parameters required by comparable models. Our source code is publicly available at https://github.com/iamtcs/GSASN.