MRM-Lasso: A Sparse Multiview Feature Selection Method via Low-Rank Analysis.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multiview data analysis is crucial for applications like image classification and video understanding.
• High-dimensional multiview data presents challenges in redundant feature removal.
• Existing multiview feature selection methods often struggle with scalability and capturing complex cross-view relationships.

Purpose of the Study:

• To propose a novel multiview feature selection algorithm, multiview rank minimization-based Lasso (MRM-Lasso).
• To enhance the performance of multiview classification by effectively selecting relevant features.
• To capture latent correlations across different views using pattern-specific weights.

Main Methods:

• Developed MRM-Lasso, combining Lasso for sparse feature selection and rank minimization for pattern learning.
• Introduced pattern-specific weights to measure sample contribution to labels within each view.
• Employed the alternating direction method of multipliers for algorithm optimization.
• Captured cross-view correlations via a low-rank matrix of pattern-specific weights.

Main Results:

• MRM-Lasso achieved superior multiview classification performance compared to baseline methods on four real-life datasets.
• The selected features by MRM-Lasso demonstrated improved classification accuracy.
• Pattern-specific weights were found to be more significant than view-specific weights for multiview data learning.

Conclusions:

• MRM-Lasso offers an effective approach for feature selection in high-dimensional multiview data.
• The proposed method successfully addresses the challenge of redundant feature removal.
• Pattern-specific weights are vital for uncovering underlying structures and improving multiview learning outcomes.