Attributed Multi-Order Graph Convolutional Network for Heterogeneous Graphs.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Mining

Background:

• Heterogeneous graph neural networks are vital for analyzing complex multi-relational data.
• Designing effective meta-paths is a key challenge impacting embedding quality in heterogeneous graph learning.
• Existing methods often struggle with automated meta-path discovery for multi-hop neighbors.

Purpose of the Study:

• To propose a novel Attributed Multi-Order Graph Convolutional Network (AMOGCN) for automated meta-path exploration.
• To enhance the learning of discriminative node embeddings and relations in heterogeneous graphs.
• To improve semi-supervised classification performance on complex graph structures.

Main Methods:

• AMOGCN aggregates multi-order adjacency matrices to explore meta-paths involving multi-hop neighbors.
• It fuses various orders of adjacency matrices into a unified multi-order adjacency matrix.
• Node semantic information, derived from attribute-based homophily, supervises the fusion process.

Main Results:

• The proposed AMOGCN model automatically discovers effective meta-paths.
• It achieves superior semi-supervised classification performance compared to state-of-the-art methods.
• The model demonstrates efficient cross-hop information propagation equivalent to multi-layer networks.

Conclusions:

• AMOGCN offers an effective approach for automated meta-path discovery in heterogeneous graphs.
• The method significantly enhances node embedding and classification accuracy.
• This work advances the field of heterogeneous graph learning by addressing key challenges in meta-path design.