Deep Multiview Clustering via Iteratively Self-Supervised Universal and Specific Space Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multiview clustering aims to group objects using information from multiple sources.
• Existing methods often rely on restrictive assumptions like linear transformations or a single latent space, limiting their effectiveness.
• These limitations hinder comprehensive object description and accurate partitioning.

Purpose of the Study:

• To develop an advanced multiview clustering technique that addresses the shortcomings of current approaches.
• To propose a model capable of learning both shared and view-specific latent spaces for improved representation.
• To enhance the collaborative representation of objects by fully exploiting cross-view relationships.

Main Methods:

• The proposed method learns common and specific sampling spaces for each view.
• It utilizes an iterative self-supervision scheme to refine the affinity matrix.
• Clustering is formulated as a convex optimization problem, solved using both linear and deep nonlinear (autoencoder) approaches.

Main Results:

• The model demonstrated superior performance across six diverse real-world datasets.
• It outperformed established benchmark methods in multiview clustering tasks.
• The learned common and specific spaces effectively captured collaborative object representations.

Conclusions:

• The proposed multiview clustering framework offers a more robust and effective solution compared to existing methods.
• Learning both common and specific latent spaces is crucial for exploiting collaborative representations.
• The approach shows significant promise for applications requiring comprehensive object analysis from multiple data sources.