Path reliability-based graph attention networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Self-attention mechanisms in Graph Neural Networks (GNNs) have advanced graph representation learning.
• Current attention-based GNNs primarily consider direct neighbors, limiting the capture of multi-hop structural information.
• This limitation impacts performance in tasks like link prediction, knowledge graph completion, and adversarial attacks.

Purpose of the Study:

• To propose Path Reliability-based Graph Attention Networks (PRGATs) for improved graph representation learning.
• To incorporate multi-hop neighboring context into attention score computation within a single layer.
• To enhance the capture of longer-range dependencies and large-scale structural information.

Main Methods:

• Introduction of Path Reliability-based Graph Attention Networks (PRGATs).
• Development of a path reliability-based attention layer utilizing a resource-constrained allocation algorithm.
• Computation of reliable paths and attention scores between neighboring and non-neighboring nodes to expand the receptive field.

Main Results:

• PRGATs demonstrated superior performance compared to baseline methods on real-world datasets.
• Achieved up to a 3% improvement in standard node classification tasks.
• Showcased a significant 12% improvement in graph universal adversarial attack scenarios.

Conclusions:

• PRGATs effectively integrate multi-hop contextual information into GNNs.
• The proposed method enhances the model's ability to capture complex graph structures and dependencies.
• PRGATs offer a promising advancement for various graph-based machine learning applications.