Self-Supervised Information Bottleneck for Deep Multi-View Subspace Clustering.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multi-view subspace clustering aims to group data by leveraging information from multiple sources.
• Existing methods often struggle to effectively capture both common and view-specific information across diverse datasets.

Purpose of the Study:

• To develop a novel deep multi-view subspace clustering framework using an information-theoretic approach.
• To introduce a self-supervised method that learns common latent representations across multiple views.
• To enhance clustering performance by explicitly modeling view-specific information.

Main Methods:

• The proposed framework, Self-supervised Information Bottleneck based Multi-view Subspace Clustering (SIB-MSC), extends the information bottleneck principle.
• SIB-MSC learns a shared latent space by maximizing common information and minimizing redundant information across views.
• Mutual information-based regularization terms are employed to disentangle view-specific latent spaces.

Main Results:

• Extensive experiments on real-world multi-view datasets were conducted.
• The SIB-MSC framework demonstrated superior performance compared to existing state-of-the-art methods.
• The method effectively captures commonalities and distinctiveness across different data views.

Conclusions:

• The proposed SIB-MSC framework offers an effective approach for deep multi-view subspace clustering.
• The information-theoretic perspective and self-supervised learning significantly improve clustering accuracy.
• SIB-MSC provides a robust method for leveraging multi-view data in clustering tasks.