BHGNN-RT: Capturing bidirectionality and network heterogeneity in graphs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph neural networks (GNNs) excel at representation learning on graph data.
• Existing GNNs struggle with directed heterogeneous graphs, limiting their application.
• Oversmoothing is a common challenge in deep GNN architectures.

Purpose of the Study:

• To propose a novel embedding method, BHGNN-RT, for directed heterogeneous graphs.
• To enhance representation learning by capturing bidirectional message flows and network heterogeneity.
• To mitigate oversmoothing in deep GNNs using a random teleportation mechanism.

Main Methods:

• Developed a bidirectional heterogeneous graph neural network with random teleport (BHGNN-RT).
• Incorporated relation-specific transformations to integrate heterogeneous edge types.
• Implemented a teleportation mechanism to address oversmoothing and improve information flow.

Main Results:

• BHGNN-RT demonstrated superior performance over state-of-the-art baselines on various datasets.
• Achieved up to 11.5% improvement in classification accuracy and 19.3% in entity clustering.
• Optimizing message components, model layers, and teleportation proportion further boosted performance.

Conclusions:

• BHGNN-RT effectively captures structural and directional information in directed heterogeneous graphs.
• The proposed method offers a robust and efficient solution for representation learning on complex graph data.
• BHGNN-RT represents a significant advancement in handling directed heterogeneous graph neural networks.