Revisiting multi-view learning: A perspective of implicitly heterogeneous Graph Convolutional Network.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Multi-view Learning

Background:

• Graph Convolutional Networks (GCNs) excel at processing single-view graph data.
• Many real-world datasets are inherently multi-view, posing challenges for traditional GCNs.
• Existing GCNs struggle with the heterophily property common in multi-view data.

Purpose of the Study:

• To propose an implicit heterogeneous graph convolutional network for multi-view data.
• To effectively capture data heterogeneity and leverage GCN's feature aggregation.
• To improve performance on tasks involving complex, multi-view datasets.

Main Methods:

• Developed an implicit heterogeneous graph convolutional network.
• Automatically assigned optimal importance to each view during meta-path graph construction.
• Explored high-order cross-view meta-paths and generated graph matrices.
• Integrated graph matrices with learnable global feature representations for multi-level embeddings.
• Introduced a graph-level attention mechanism for individual node information allocation.

Main Results:

• The proposed method effectively captures heterogeneity in multi-view data.
• Achieved superior performance compared to state-of-the-art approaches in extensive experiments.
• Demonstrated the capability to leverage both local and global information through attention mechanisms.

Conclusions:

• The implicit heterogeneous graph convolutional network is a powerful tool for multi-view learning.
• The method offers a significant advancement in handling complex, multi-view graph data.
• The approach provides a robust framework for feature aggregation and information utilization in heterogeneous graph environments.