Learning Disentangled Representations for Generalized Multi-View Clustering

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Multi-View Clustering (MVC) leverages diverse data perspectives.
• Deep MVC methods face challenges with view-distribution entanglement during fusion.
• This entanglement degrades latent space quality and clustering outcomes.

Purpose of the Study:

• To introduce a novel framework, Generalized Multi-view Auto-Encoder (GMAE), for effective MVC.
• To address view-distribution entanglement and preserve cross-view complementarity.
• To enhance the quality of shared latent space for superior clustering.

Main Methods:

• GMAE employs dual-path autoencoders for disentangled representation learning.
• Source features are decoupled into view-specific and view-common embeddings.
• Cross-view adversarial discriminators guide feature extraction and mutual information is modulated to align distributions.

Main Results:

• GMAE successfully decouples features, revealing clearer clustering structures.
• The framework generates robust, non-trivial embeddings by preventing representation collapse.
• Experiments on 13 benchmark datasets show GMAE outperforms state-of-the-art methods.

Conclusions:

• GMAE offers a significant advancement in deep Multi-View Clustering.
• The proposed method demonstrates superior performance in both complete and incomplete MVC tasks.
• The framework effectively preserves cross-view complementarity for improved clustering.