Multiview Subspace Clustering With Multilevel Representations and Adversarial Regularization.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multiview subspace clustering is effective for uncovering data structure.
• Existing methods often neglect nonlinearities, multilevel representations, and latent data distributions.

Purpose of the Study:

• To propose a novel Multiview Subspace Clustering with MLRs and Adversarial Regularization (MvSC-MRAR) method.
• To address limitations in current multiview clustering techniques.

Main Methods:

• Utilizes deep auto-encoders for nonlinear structure modeling.
• Incorporates self-expressive layers for multilevel latent representations.
• Employs adversarial training with a universal discriminator for realistic affinity matrices.

Main Results:

• MvSC-MRAR demonstrates significant improvements over state-of-the-art methods.
• Achieves higher clustering accuracy (ACC) and normalized mutual information (NMI).
• Validated on nine real-world multiview datasets.

Conclusions:

• The proposed MvSC-MRAR effectively integrates nonlinear structure, multilevel representations, and adversarial regularization.
• Offers a more robust approach to multiview subspace clustering.
• Outperforms existing methods on benchmark datasets.