Non-Negative Spectral Learning and Sparse Regression-Based Dual-Graph Regularized Feature Selection.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• High-dimensional data analysis often suffers from the curse of dimensionality.
• Existing feature selection methods primarily focus on data space, neglecting valuable feature space information.
• This limitation hinders the full utilization of data characteristics and can lead to suboptimal feature selection.

Purpose of the Study:

• To propose a novel unsupervised feature selection algorithm, NSSRD, that leverages information from both data and feature spaces.
• To enhance feature selection by integrating joint embedding learning with sparse regression and dual-graph regularization.
• To address the limitations of existing methods by incorporating non-negative constraints and L1-norm regularization.

Main Methods:

• Developed NSSRD, a feature selection algorithm based on joint embedding learning and sparse regression.
• Introduced a feature graph to exploit geometric information in the feature space alongside the data space.
• Employed non-negative constraints on embedding matrices and L1-norm regularization for sparse transformation matrix optimization.
• Utilized an iterative and alternative updating rule for efficient objective function optimization.

Main Results:

• NSSRD demonstrated superior performance compared to existing feature selection algorithms across various datasets.
• The algorithm effectively reduced data dimensionality while preserving essential discriminative information.
• Simultaneous exploitation of data and feature space geometric information proved beneficial for feature selection accuracy.

Conclusions:

• NSSRD offers a significant advancement in unsupervised feature selection for high-dimensional data.
• The proposed method effectively integrates spectral learning, sparse regression, and dual-graph regularization.
• Experimental results validate the efficiency and effectiveness of NSSRD in selecting representative features.