Graph structure learning layer and its graph convolution clustering application.

Area of Science:

• Graph representation learning
• Deep learning for graph data
• Network science

Background:

• Existing graph structure learning methods often fail with noisy or outlier-corrupted data.
• The two-step paradigm of constructing graph structure then message passing is vulnerable to unreliable learned structures.

Purpose of the Study:

• To develop a robust graph convolutional clustering network that can learn accurate graph embeddings from noisy data.
• To address the limitations of traditional two-step approaches in graph structure learning.

Main Methods:

• Proposes an adaptive graph convolutional clustering network with layer-by-layer adjustment of graph structure and node representations.
• Introduces a Graph Structure Learning layer utilizing an optimal self-expression problem solved via an efficient iterative optimization algorithm.
• Integrates an optimization process as a novel Graph Network layer.

Main Results:

• The proposed method demonstrates effectiveness in defending against the negative impacts of inaccurate graph structures.
• Experimental results validate the robustness and improved performance of the adaptive network.

Conclusions:

• The adaptive graph convolutional clustering network offers a more reliable approach to learning graph embeddings from corrupted data.
• Integrating optimization processes within network layers represents a novel direction in deep learning for graph data.