Graph Influence Network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) excel at pattern extraction but often fail to discriminate nodes when selecting information sources.
• Existing GNNs either treat all neighbors equally or distinguish them based solely on graph structure or node features.

Purpose of the Study:

• To propose a novel GNN framework, the graph influence network (GINN), that effectively discriminates neighbors by evaluating their influence on target nodes.
• To enhance graph representation learning by incorporating both structural and feature-based influence assessments for selecting information sources.

Main Methods:

• Introduced the concept of an 'Influence Set' to identify influential neighbors.
• Utilized local graph structure to construct the influence set, employing a HodgeRank-based algorithm to estimate structural influences.
• Applied an attention mechanism to measure feature influences within the influence set, filtering task-irrelevant nodes.

Main Results:

• The proposed GINN framework successfully identifies influential nodes by considering both topological structures and node features.
• Experiments demonstrated that GINN achieves state-of-the-art performance compared to existing baseline models.
• The results validate the effectiveness of discriminating neighbors in graph representation learning.

Conclusions:

• GINN offers a significant advancement in GNNs by enabling a more nuanced selection of information sources.
• The framework's ability to assess and utilize neighbor influence leads to improved graph representation learning.
• Discriminating neighbors based on combined structural and feature influences is crucial for superior GNN performance.