Domain-Adaptive Graph Attention-Supervised Network for Cross-Network Edge Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph neural networks (GNNs) struggle with noisy edges that connect dissimilar nodes, degrading performance.
• Existing GNNs address noisy edges within a single network but not across different networks.

Purpose of the Study:

• To address the novel problem of cross-network homophilous and heterophilous edge classification (CNHHEC).
• To propose a domain-adaptive graph attention-supervised network (DGASN) for effective CNHHEC.

Main Methods:

• DGASN utilizes a multihead graph attention network (GAT) encoder for joint node and edge embedding training.
• It employs direct supervision on graph attention learning using source network edge labels.
• Adversarial domain adaptation is used to minimize domain divergence and facilitate knowledge transfer.

Main Results:

• DGASN effectively distinguishes homophilous from heterophilous edges through label-discriminative embeddings.
• The method reduces negative impacts of heterophilous edges and enhances positive impacts of homophilous edges.
• State-of-the-art performance in CNHHEC was achieved on real-world datasets.

Conclusions:

• DGASN offers a pioneering solution for cross-network edge classification.
• The proposed method enhances GNN robustness against noisy edges in multi-network scenarios.
• DGASN demonstrates significant improvements in handling both homophilous and heterophilous edges across domains.