Improved Dual Correlation Reduction Network With Affinity Recovery.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Deep graph clustering aims to partition nodes without supervision.
• Existing methods often suffer from representation collapse, limiting discriminative power.
• This leads to suboptimal clustering performance due to poor latent embeddings.

Purpose of the Study:

• To propose a novel deep graph clustering algorithm, the Improved Dual Correlation Reduction Network (IDCRN).
• To address the representation collapse problem and enhance the discriminative capability of nodes.
• To improve clustering performance by refining latent representations and feature discrimination.

Main Methods:

• IDCRN approximates the cross-view feature correlation matrix to an identity matrix, reducing feature redundancy.
• It forces the cross-view sample correlation matrix to approximate a clustering-refined adjacency matrix.
• A propagation regularization term is introduced to mitigate oversmoothing in graph convolutional networks (GCNs).

Main Results:

• IDCRN explicitly improves the discriminative capability of the latent space by reducing feature dimension redundancy.
• It implicitly enhances feature discrimination by guiding latent representations to recover affinity matrices across views.
• Experiments on six benchmarks show IDCRN outperforms state-of-the-art deep graph clustering algorithms in effectiveness and efficiency.

Conclusions:

• IDCRN offers an effective solution to representation collapse in deep graph clustering.
• The proposed methods enhance both explicit and implicit discriminative capabilities for improved clustering.
• IDCRN demonstrates superior performance and efficiency compared to existing methods.