Pseudo-Label Guided Collective Matrix Factorization for Multiview Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Real-world data often contains multiple features or views, necessitating multiview clustering techniques.
• Existing methods face scalability challenges with large datasets and suboptimal results due to separate learning of representations and clusters.

Purpose of the Study:

• To propose a novel Pseudo-Label guided Collective Matrix Factorization (PLCMF) method.
• To address the limitations of existing multiview clustering methods regarding scalability and joint learning.

Main Methods:

• PLCMF utilizes pseudo-labels from individual view clustering to guide collective matrix factorization.
• It learns unified latent representations preserving both intra-view and inter-view similarities.
• A joint framework integrates representation and cluster structure learning with adaptive view weighting.

Main Results:

• The proposed PLCMF method demonstrates superior clustering accuracy compared to state-of-the-art methods.
• PLCMF achieves high computational efficiency with linear complexity, making it scalable to large datasets.
• Experiments on six benchmark datasets validate the effectiveness of the PLCMF approach.

Conclusions:

• PLCMF offers an effective and efficient solution for multiview clustering.
• The joint learning framework and pseudo-label guidance overcome limitations of traditional two-step approaches.
• The method shows significant improvements in both clustering performance and computational speed.