Edge-Wise Graph-Instructed Neural Networks

2024-09-12

Journal of Computer Science

AI Summary

Graph Neural Networks (GNNs) suffer from inaccurate predictions and training instability on synthetic graphs (e.g., Barabási–Albert and Erdős–Rényi), while the node-centric paradigm inherently limits fine-grained modeling of edge-level relational semantics. Method: We propose the Edge-level Graph Instruction Neural Network (EGINN), introducing the first edge-level graph instruction paradigm. EGINN employs an instruction embedding encoder to generate edge-heterogeneity-aware dynamic instructions, integrated with graph attention and an edge-conditioned gating unit to enable differentiable, interpretable message routing and feature transformation—departing from static edge aggregation and node-centric constraints in conventional GNNs. Contribution/Results: Evaluated on six benchmark graph learning tasks, EGINN achieves an average accuracy gain of 2.7%, significantly improves edge-level reasoning interpretability, and demonstrates superior generalization and robustness over state-of-the-art GNNs.

Technology Category

Machine Learning: Graph-based Machine LearningData Mining & Knowledge Management: Graph Mining, Social Network Analysis & CommunityNatural Language Processing: Sentence-level Semantics, Textual Inference, etc.

Application Category

Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphsSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsEconomics, Online Markets and Human Computation: Incentives in network design for Web infrastructures and ecosystems

Problem

• Graph Neural Networks
• Prediction Accuracy
• Training Stability

Innovation

• Edge-Wise Guided GI layer
• Graph Inductive Neural Network
• Training Stability