Multiview Deep Subspace Clustering Networks.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multiview subspace clustering integrates complementary data information.
• Existing methods often rely on handcrafted features and separate learning stages.
• Limitations include unintegrated multiview relations and incompatibility with deep learning's end-to-end nature.

Purpose of the Study:

• To propose a novel end-to-end deep learning framework for multiview subspace clustering.
• To address limitations of traditional methods by embedding multiview relations into feature learning.
• To develop a flexible network architecture adaptable to various datasets.

Main Methods:

• Introduced Multiview Deep Subspace Clustering Networks (MvDSCNs) with diversity (Dnet) and universality (Unet) subnetworks.
• Utilized deep convolutional autoencoders to build a latent space for self-representation matrix learning.
• Incorporated Hilbert-Schmidt Independence Criterion (HSIC) for diversity regularization and universality regularization for alignment.

Main Results:

• MvDSCNs effectively learn view-specific and common self-representation matrices in an end-to-end manner.
• The Hilbert-Schmidt Independence Criterion captured nonlinear, high-order multiview relations.
• The framework demonstrated superior clustering performance across experiments, unifying multiple backbones.

Conclusions:

• MvDSCNs offer a robust and flexible approach to multiview subspace clustering.
• The proposed method effectively leverages complementary information from multiple views.
• The end-to-end learning and unified backbone approach significantly advance the field.