Self-Supervised Deep Multiview Spectral Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multiview spectral clustering is a powerful unsupervised learning technique.
• Pairwise constraints significantly enhance clustering but typically require manual annotation.
• Manual annotation is resource-intensive and limits scalability.

Purpose of the Study:

• To develop a self-supervised multiview spectral clustering method.
• To reduce reliance on human-labeled constraints.
• To improve clustering performance by automatically generating pairwise constraints.

Main Methods:

• Utilized fused multiple autoencoders for latent consistent feature extraction across views.
• Generated pairwise constraints based on inter-view commonality.
• Employed neural network with historical memory for constraint propagation and optimization of the fused affinity matrix.

Main Results:

• Demonstrated the effectiveness of the proposed self-supervised approach.
• Achieved improved multiview spectral clustering performance compared to existing methods.
• Validated through experiments on four benchmark datasets.

Conclusions:

• The proposed self-supervised multiview spectral clustering effectively leverages automatically retrieved constraints.
• This method alleviates the need for manual annotation, saving human resources.
• The approach shows significant potential for enhancing unsupervised multiview clustering tasks.