Composite attention mechanism network for deep contrastive multi-view clustering.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Deep multi-view clustering methods commonly use autoencoders and contrastive learning.
• Existing methods often fail to capture inherent relationships within or across views.
• Limited contrastive modules can hinder communication between representations and clustering assignments.

Purpose of the Study:

• To propose a new composite attention framework for deep multi-view clustering.
• To address limitations in representation learning and feature communication in existing methods.
• To improve clustering performance on unlabeled multi-view data.

Main Methods:

• Utilizing a composite attention structure with Hierarchical Transformers and Shared Attention for representation learning.
• Implementing a dual contrastive framework with a novel communication loss.
• Enhancing latent representations to preserve intra-view features and balance cross-view contributions.

Main Results:

• The proposed method achieves superior clustering performance compared to state-of-the-art algorithms.
• Significant accuracy improvements, averaging 10%, were observed on the Caltech dataset and its subsets.
• The framework effectively mines inherent relationships and enhances communication between representations and cluster assignments.

Conclusions:

• The composite attention framework offers a more effective approach to deep multi-view clustering.
• The method demonstrates robust performance in handling unlabeled multi-view data, particularly for segmentation tasks.
• This work advances the field by improving representation learning and feature integration in clustering.