Unsupervised Feature Selection for High-Order Embedding Learning and Sparse Learning.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Unsupervised feature selection methods often overlook high-order data similarity, leading to suboptimal similarity graphs.
• High complexity and computational cost limit the applicability of existing methods, especially for high-dimensional data.

Purpose of the Study:

• To propose a novel unsupervised feature selection method, UFSHS, that addresses limitations of existing approaches.
• To leverage high-order similarity for constructing accurate data representations and selecting optimal feature subsets.

Main Methods:

• UFSHS utilizes high-order data similarity to construct an optimal similarity graph, capturing the intrinsic geometric structure.
• A unified framework integrating high-order embedding and sparse learning is employed to learn a row-sparse projection matrix.
• A novel alternative optimization strategy is developed, adapting to data dimensionality and instance count to reduce computational complexity.

Main Results:

• The proposed UFSHS method effectively selects optimal feature subsets by integrating high-order embedding and sparse learning.
• The alternative optimization strategy significantly reduces computational complexity and demonstrates applicability to various models like ridge regression, broad learning, and fuzzy systems.
• Extensive experiments on nine public datasets validate the superiority and efficiency of UFSHS compared to existing methods.

Conclusions:

• UFSHS offers a superior and efficient approach to unsupervised feature selection for high-dimensional data.
• The method's ability to capture high-order similarities and its adaptable optimization strategy enhance its practical applicability.
• UFSHS provides a robust framework for feature selection with potential extensions to other machine learning models.