Path-aware multi-scale learning for heterogeneous graph neural network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Heterogeneous Graph Neural Networks (HGNNs) model complex data with diverse node/edge types.
• Meta-path-based HGNNs offer performance and interpretability but often ignore meta-path redundancy and global information.
• Existing HGNNs struggle with comprehensive representation learning due to limited utilization of path characteristics and global context.

Purpose of the Study:

• To propose PM-HGNN, a path-aware multi-scale heterogeneous graph neural network.
• To address limitations in meta-path redundancy and insufficient global information utilization in current HGNNs.
• To enhance representation learning for heterogeneous graphs.

Main Methods:

• PM-HGNN employs a global similarity-based mean aggregator for pre-computing neighbor aggregation.
• It dynamically assigns weights to meta-paths, exploiting their relevance and differences for redundancy reduction.
• The model integrates multi-scale information and path characteristics for improved learning.

Main Results:

• PM-HGNN consistently outperforms state-of-the-art methods on node classification tasks.
• Experiments on four real-world heterogeneous graph datasets validate the proposed approach.
• The method demonstrates superior performance in capturing complex graph structures.

Conclusions:

• PM-HGNN effectively addresses the limitations of existing meta-path-based HGNNs.
• The proposed approach enhances representation learning by incorporating global information and optimized meta-path utilization.
• PM-HGNN represents a significant advancement in heterogeneous graph neural network research.