Attention-based stackable graph convolutional network for multi-view learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Graph Convolutional Networks (GCNs) are effective for multi-view learning but face challenges with complex preprocessing and over-smoothing.
• Existing GCN methods often require computationally expensive sparsification and suffer performance degradation with increased network depth.

Purpose of the Study:

• To propose an attention-based stackable GCN to mitigate over-smoothing and enhance multi-view learning.
• To address computational costs and training difficulties associated with traditional GCN methods.

Main Methods:

• Introduced node self-attention for dynamic node connections and view-specific representations.
• Developed a data-driven approach for attention-based view weighting to ensure cross-view consistency.
• Integrated an attention mechanism with residual connectivity to compensate for information loss during graph convolution.

Main Results:

• The proposed attention-based GCN effectively captures cross-view consistency and mitigates over-smoothing.
• Node self-attention dynamically establishes connections, improving representation generation.
• Experimental results show superior performance compared to state-of-the-art methods in multi-view semi-supervised tasks.

Conclusions:

• The attention-based stackable GCN offers a robust solution for multi-view learning challenges.
• The method enhances GCN capabilities by effectively managing over-smoothing and information loss.
• This approach demonstrates significant improvements in multi-view semi-supervised learning performance.