Linear Complexity Multi-View Unsupervised Feature Selection via Anchor-Based Feature Relationship Construction.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Multi-view unsupervised feature selection methods often focus on sample relationships, neglecting crucial feature relationships.
• Existing methods face high computational complexity (O(d^2) or higher) when constructing complete feature graphs.
• There is a need for efficient methods to extract feature importance directly from feature graphs.

Purpose of the Study:

• To develop a novel multi-view unsupervised feature selection algorithm with reduced complexity.
• To introduce an anchor-based strategy and feature bipartite graphs for efficient feature relationship modeling.
• To design a low-complexity method for direct feature importance extraction from feature bipartite graphs.

Main Methods:

• An anchor-based strategy and feature bipartite graph construction are employed to reduce complexity.
• A novel method is proposed to directly obtain feature scores from the feature bipartite graph, reducing time complexity from O(d^3) to O(d).
• Self-expressive multi-view subspace learning adaptively learns feature-level anchor graph structures, capturing feature-anchor relationships and multi-view consistency.

Main Results:

• The proposed method achieves O(nd) space and time complexity, a significant improvement over existing approaches.
• Experimental results on image and biological datasets demonstrate the superiority of the proposed algorithm compared to seven state-of-the-art methods.
• The method effectively captures structural information between features and anchors, as well as consistency and complementary information across views.

Conclusions:

• The proposed algorithm offers an efficient and effective solution for multi-view unsupervised feature selection.
• The anchor-based strategy and feature bipartite graph approach significantly reduce computational complexity.
• This work represents a novel contribution to the field of multi-view unsupervised feature selection with practical implications for handling large datasets.