Influence problems on a content-spreading model and graph machine learning

Area of Science:

• Network Science
• Computational Social Science
• Machine Learning on Graphs

Background:

• Modeling information or disease spread on networks is a key challenge in network science.
• Influence maximization aims to find seed nodes that maximize spread, while influence computation seeks exact influence probabilities.
• Existing models like the independent cascade model are well-studied but may not capture all real-world dynamics.

Purpose of the Study:

• To introduce and analyze a new content-spreading model inspired by bounded confidence models.
• To generalize the independent cascade model and propose novel influence computation and maximization problems.
• To evaluate graph neural networks (GNNs) for predicting influence probabilities and assessing the oversquashing problem.

Main Methods:

• Introduced a novel content-spreading model based on bounded confidence.
• Demonstrated the generalization of the independent cascade model.
• Developed and applied centrality measures for influence identification in tree networks.
• Trained graph neural networks to predict influence probabilities.

Main Results:

• The proposed content-spreading model encompasses the independent cascade model as a special case.
• Analytical and computational tractability was achieved for influence problems on tree networks.
• Graph neural networks show promise as a benchmark for evaluating influence prediction and the oversquashing phenomenon.

Conclusions:

• The new bounded confidence-inspired model offers a broader framework for studying network contagion.
• Centrality measures provide efficient solutions for influence problems on tree structures.
• GNNs present a valuable tool for advancing research in network influence modeling and GNN architecture evaluation.