Enhanced anchor contrastive multi-view representations learning network for clustering.

Area of Science:

• Machine Learning
• Computer Vision
• Data Mining

Background:

• Deep multi-view contrastive clustering excels at sample classification and information consistency.
• Existing models face challenges with high complexity for large-scale data, inaccurate anchor points from K-means, false-negative pairs in instance-level contrastive learning (CL), and ignoring cluster-level information.

Purpose of the Study:

• To propose a novel Enhanced Anchor Contrastive Multi-view Representations Learning Network for Clustering (EACMVC).
• To address limitations in existing deep multi-view contrastive clustering models, including complexity, anchor point accuracy, and discriminative structure learning.

Main Methods:

• Introduced anchor representations to reduce model complexity and improve scalability.
• Developed a global structure-guided anchor representations CL module (GSgARCL) to mitigate false-negative pairs.
• Implemented a self-supervised label alignment module (SsLA) for enhanced feature representation and accurate clustering structure learning using Kullback-Leibler (KL) divergence.

Main Results:

• The proposed EACMVC model demonstrates superior performance compared to state-of-the-art algorithms.
• The integration of anchor representations, GSgARCL, and SsLA modules proved effective in enhancing clustering accuracy and efficiency.
• The model successfully learned target distributions and aligned them with soft labels for improved clustering structures.

Conclusions:

• EACMVC effectively overcomes the limitations of existing deep multi-view contrastive clustering methods.
• The complementary and mutually supportive modules enhance the model's ability to handle large-scale data and achieve more discriminative clustering.
• The proposed approach offers a robust and efficient solution for multi-view clustering tasks.