Deep Multiview Clustering by Pseudo-Label Guided Contrastive Learning and Dual Correlation Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep multiview clustering (MVC) leverages relationships across multiple data views for improved performance.
• Existing deep MVC methods face challenges with inaccurate contrastive learning (CL) due to improper pair selection and limited exploration of dual correlations (feature and cluster).

Purpose of the Study:

• To propose a novel deep MVC framework addressing the limitations of current methods.
• To enhance the accuracy of contrastive learning in MVC by refining positive and negative pair selection.
• To explore both feature and cluster correlations across views for more comprehensive relation learning.

Main Methods:

• Developed a pseudo-label guided contrastive learning (CL) mechanism to accurately align feature distributions by mitigating false negative pairs.
• Introduced dual correlation learning to investigate both feature and cluster correlations among multiple views.
• Implemented a novel deep MVC framework integrating these two components.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art deep MVC techniques across various datasets.
• Pseudo-label guided CL effectively improves feature distribution alignment and discriminative feature learning.
• Dual correlation learning captures richer and more comprehensive relationships across views.

Conclusions:

• The novel deep MVC framework effectively addresses key challenges in existing methods, leading to enhanced clustering performance.
• Pseudo-label guided CL and dual correlation learning are crucial for accurate and comprehensive multiview representation learning.
• The proposed approach offers a significant advancement in the field of deep multiview clustering.