Low-Rank Tensor Regularized Views Recovery for Incomplete Multiview Clustering.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Real-world multiview data frequently exhibit missing views, posing challenges for existing incomplete multiview clustering (IMVC) methods.
• Current IMVC techniques often fail to fully exploit information from missing data or capture inter-view correlations effectively.

Purpose of the Study:

• To propose a novel method, Low-rank Tensor regularized viEws Recovery (LATER), for incomplete multiview clustering.
• To jointly reconstruct and utilize missing views while learning multilevel graphs for comprehensive similarity discovery.

Main Methods:

• LATER recovers missing views from a common latent representation, with recovered views enhancing shared pattern learning.
• Multilevel graphs are learned via self-representation using shared subspace representations and recovered complete multiview data.
• A tensor nuclear norm regularizer is employed to enforce global low-rank properties and explore inter-view correlations.
• An alternating direction minimization algorithm optimizes the model, complemented by a new initialization strategy.

Main Results:

• The proposed LATER method demonstrates superior performance compared to state-of-the-art approaches in extensive experiments.
• The joint reconstruction and utilization of missing views lead to more robust and accurate clustering outcomes.
• Learning multilevel graphs effectively captures both consistent and complementary information across views.

Conclusions:

• The LATER method offers a significant advancement in incomplete multiview clustering by effectively handling missing data.
• The unified model successfully integrates missing view recovery, multilevel graph learning, and tensor regularization for enhanced clustering.
• The proposed approach provides a promising direction for future research in multiview learning with incomplete data.