Enhancing node influence prediction in large networks via multi-Level knowledge distillation.

Area of Science:

• Network analysis
• Computational social science
• Machine learning

Background:

• Predicting node influence in large-scale complex networks is crucial but computationally expensive.
• Traditional methods like the Susceptible-Infected-Recovered (SIR) model are too slow for large networks, hindering scalability.

Purpose of the Study:

• To develop a computationally efficient method for predicting node influence in large networks.
• To enhance prediction accuracy and reduce inference time, especially when labeled data is scarce.

Main Methods:

• Utilized multi-level knowledge distillation with a teacher-student architecture.
• Implemented knowledge transfer from richly labeled networks to sparsely labeled ones.
• Designed a shallow student model with few parameters for reduced inference time.
• Incorporated soft labels and adversarial alignment for knowledge transfer.

Main Results:

• Achieved significant improvements in predictive accuracy compared to existing methods.
• Demonstrated substantial reductions in computational inference time.
• Validated the approach on various real-world network datasets.

Conclusions:

• Multi-level knowledge distillation offers an effective and scalable solution for node influence prediction.
• The proposed shallow student model significantly enhances computational efficiency.
• This approach is particularly beneficial for large-scale networks with limited labeled nodes.