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

Area of Science:

• Computer Science
• Network Engineering
• Artificial Intelligence

Background:

• Network traffic prediction is crucial for intelligent routers facing increasing data and video demands.
• Current spatial models like GCNs and GATs struggle with nonlinear spatial relationships due to linear constraints in attention mechanisms.
• Accurate prediction is vital for resource planning, bandwidth management, and latency reduction.

Purpose of the Study:

• To develop a novel network traffic prediction model that overcomes the limitations of existing spatial approaches.
• To introduce a self-learning attention mechanism that captures complex nonlinear spatial dependencies.
• To propose an integrated spatio-temporal model for enhanced prediction accuracy and efficiency.

Main Methods:

• Proposed the Graph Self-learning Attention Scores Network (GSASN) with a learnable parameter matrix to autonomously learn attention scores.
• Developed a spatio-temporal model (ST-GSASN) by integrating GSASN with temporal learning via gated fusion.
• Validated the models on two real-world network traffic datasets.

Main Results:

• GSASN demonstrated significant improvements over Graph Attention Networks (GAT), reducing RMSE by up to 24.7% and MAE by 36.5%.
• ST-GSASN outperformed baseline and state-of-the-art methods, including transformer-based models, in prediction accuracy.
• ST-GSASN achieved superior computational efficiency, requiring only 5% of the parameters of comparable models.

Conclusions:

• The proposed GSASN effectively captures complex nonlinear spatial relationships in network traffic.
• ST-GSASN offers a highly accurate and computationally efficient solution for spatio-temporal network traffic prediction.
• The novel self-learning attention mechanism provides a promising direction for future network intelligence research.