HAGNN: Hybrid Aggregation for Heterogeneous Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Heterogeneous graph neural networks (GNNs) are effective for handling complex graph data.
• Meta-paths are crucial in existing heterogeneous GNNs, but their necessity is debated.
• Meta-path-free models show comparable performance, questioning the exclusive reliance on meta-paths.

Purpose of the Study:

• To investigate the intrinsic differences between meta-path-based and meta-path-free neighbor selection in GNNs.
• To propose a novel framework, Hybrid Aggregation for Heterogeneous GNNs (HAGNN), for comprehensive semantic information utilization.
• To enhance heterogeneous GNNs by simultaneously leveraging meta-path and directly connected neighbors.

Main Methods:

• HAGNN employs a two-phase aggregation: meta-path-based intratype and meta-path-free intertype aggregation.
• A fused meta-path graph data structure is introduced for structural semantic aware aggregation.
• Embeddings from both aggregation phases are combined to capture rich graph semantics.

Main Results:

• HAGNN effectively utilizes the heterogeneity of graphs by combining different aggregation strategies.
• Experiments on node classification, clustering, and link prediction demonstrate HAGNN's superiority.
• The proposed framework shows significant improvements over existing heterogeneous GNN models.

Conclusions:

• HAGNN offers a comprehensive approach to heterogeneous graph representation learning.
• The framework effectively integrates diverse semantic information, outperforming existing methods.
• HAGNN demonstrates enhanced effectiveness and efficiency in various graph-based tasks.