Disentangled representation learning for multi-view clustering via von Mises-Fisher hyperspherical embedding.

Area of Science:

• Machine Learning
• Data Science
• Computational Statistics

Background:

• Multi-view clustering integrates diverse data but often assumes Gaussian latent spaces, limiting performance on complex, high-dimensional data.
• Existing methods struggle with data alignment, information fusion, and similarity measurement due to non-Gaussian distributions.

Purpose of the Study:

• To propose a novel contrastive multi-view clustering framework addressing limitations of Gaussian assumptions in latent spaces.
• To improve data alignment, information fusion, and similarity measurement for complex datasets.

Main Methods:

• Developed a framework using hyperspherical embeddings modeled by the von Mises-Fisher (vMF) distribution.
• Incorporated a contrastive learning paradigm with alignment and uniformity losses for discriminative representations.
• Optimized intra-cluster cohesion and inter-cluster separability across multiple views.

Main Results:

• The proposed method significantly outperforms state-of-the-art approaches on benchmark datasets.
• Demonstrated superior performance particularly with high-dimensional and complex, non-Gaussian data distributions.
• Achieved enhanced intra-cluster cohesion and inter-cluster separability.

Conclusions:

• The hyperspherical embedding approach with vMF distribution and contrastive learning effectively handles complex multi-view data.
• The framework offers a robust solution for multi-view clustering challenges, improving representation learning and clustering accuracy.
• Publicly available code facilitates reproducibility and further research.