Hierarchical knowledge distillation framework for efficient node influence prediction in large-scale complex...

Area of Science:

• Network Science
• Computational Social Science
• Machine Learning

Background:

• Node influence prediction is crucial for applications like epidemic control and viral marketing.
• Traditional Susceptible-Infected-Recovered (SIR) simulations are computationally intensive, hindering real-time analysis on large networks.
• Existing methods struggle to balance accuracy and computational efficiency.

Purpose of the Study:

• To develop a computationally efficient framework for accurate node influence prediction on large-scale networks.
• To reduce the computational cost and simulation requirements of traditional methods.
• To enable real-time node influence prediction.

Main Methods:

• Proposed HKD-NIP (Hierarchical Knowledge Distillation for Node Influence Prediction) framework.
• Utilized a dual-teacher architecture with a general teacher (trained on diverse synthetic networks) and a domain-specific teacher.
• Employed a LightGCN-based student model for knowledge distillation via soft labels and contrastive representation alignment.
• Implemented a hierarchical two-stage distillation process to bridge the structural domain gap.

Main Results:

• Achieved simulation-level accuracy in node influence prediction.
• Reduced computational time by 89% and SIR simulation requirements by 90%.
• Demonstrated superior performance on real-world datasets with Kendall's τ of 0.921 (15.4% improvement) and MSE of 0.0085 (46% improvement).
• Showcased scalability on networks up to 500,000 nodes, with sub-second inference times.

Conclusions:

• HKD-NIP effectively balances computational efficiency and prediction accuracy for real-time deployment.
• The hierarchical distillation approach enables the model to capture both universal and network-specific propagation patterns.
• The framework significantly outperforms state-of-the-art methods in node influence prediction tasks.