Autoweighted Multiview Feature Selection With Graph Optimization.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Unsupervised multiview feature selection is crucial for high-dimensional data in multiview learning.
• Existing graph-based methods often overlook cross-view data structure and are sensitive to noise.
• Predefined Laplacian graphs limit optimal neighbor assignment in current approaches.

Purpose of the Study:

• To propose a novel unsupervised multiview feature selection model based on graph learning.
• To address limitations of existing methods by learning a consensus similarity graph and incorporating adaptive view weighting.
• To improve feature selection accuracy and robustness in high-dimensional multiview data.

Main Methods:

• Learning a consensus similarity graph shared across different views during feature selection.
• Incorporating a rank constraint to optimize the similarity matrix for enhanced accuracy.
• Developing an auto-weighted framework for adaptive view weight assignment.
• Utilizing an effective alternative iterative algorithm for problem optimization.

Main Results:

• The proposed model reveals underlying data relationships from selected feature subsets.
• Optimized similarity matrices provide more accurate information.
• Adaptive view weighting improves the model's ability to handle diverse data.
• Experimental results demonstrate superior performance over state-of-the-art methods on various datasets.

Conclusions:

• The novel graph learning-based approach significantly enhances unsupervised multiview feature selection.
• The method's ability to learn consensus graphs and adapt view weights offers improved robustness and accuracy.
• This work provides a more effective solution for feature selection in high-dimensional multiview learning scenarios.