Modal-Regression-Based Structured Low-Rank Matrix Recovery for Multiview Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Low-rank Multiview Subspace Learning (LMvSL) shows promise for cross-view classification.
• Existing LMvSL methods struggle with simultaneous view discrepancy and discriminancy, degrading performance on heterogeneous data.
• Handling complex noise distributions beyond Gaussian or Laplacian is a practical challenge in low-rank modeling.

Purpose of the Study:

• To propose a novel method, Structured Low-rank Matrix Recovery (SLMR), to effectively mitigate view discrepancy and enhance discriminancy in multiview data.
• To develop an advanced framework, MR-SLMR, by integrating modal regression into SLMR to robustly handle diverse and unknown noise types.
• To demonstrate the superiority and noise robustness of the proposed MR-SLMR method through empirical evaluation.

Main Methods:

• Proposed Structured Low-rank Matrix Recovery (SLMR) based on block-diagonal representation learning to recover a structured low-rank matrix.
• Integrated modal regression into SLMR, creating MR-SLMR, to address arbitrary zero-mean noise variables including outliers.
• Employed the Alternating Direction Method of Multipliers (ADMM) framework and half-quadratic theory for efficient MR-SLMR optimization.

Main Results:

• MR-SLMR effectively removes view discrepancy and improves data discriminancy.
• The proposed method demonstrates robustness against various complex noise types, including Gaussian noise, random noise, and outliers.
• Experimental results on four public databases confirm the superior performance of MR-SLMR compared to existing LMvSL methods.

Conclusions:

• MR-SLMR offers a significant advancement in Low-rank Multiview Subspace Learning by effectively handling view discrepancy and complex noise.
• The method's ability to manage diverse noise distributions makes it practical for real-world applications.
• MR-SLMR establishes a new state-of-the-art in cross-view classification tasks with heterogeneous data.