Graph Transformer Networks: Learning meta-path graphs to improve GNNs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Graph Neural Networks (GNNs) excel at learning from graph-structured data.
• Existing GNNs struggle with fixed, homogeneous graphs, misspecified graphs, and heterogeneous data.
• Learning effective node representations on dynamic or complex graph structures remains a challenge.

Purpose of the Study:

• To introduce Graph Transformer Networks (GTNs) for generating optimal graph structures.
• To enhance GTNs with Fast Graph Transformer Networks (FastGTNs) for improved scalability.
• To extend graph transformations to capture semantic proximity beyond traditional meta-paths.

Main Methods:

• Proposing Graph Transformer Networks (GTNs) to dynamically generate graph structures.
• Developing Fast Graph Transformer Networks (FastGTNs) for efficient graph transformation and scalability.
• Incorporating non-local operations based on semantic proximity for enhanced node feature learning.

Main Results:

• GTNs and FastGTNs achieve state-of-the-art performance in node classification tasks.
• FastGTNs demonstrate significant improvements in training and inference speed (up to 230x faster) and memory usage (up to 100x less) compared to GTNs.
• Non-local operations extend graph transformations beyond meta-paths, improving representation quality.

Conclusions:

• GTNs and FastGTNs effectively address limitations of traditional GNNs on complex graph data.
• FastGTNs provide a scalable and efficient solution for graph transformation and representation learning.
• The proposed methods achieve superior performance on node classification across homogeneous and heterogeneous graphs.