Feature Selection Based on a Sparse Neural-Network Layer With Normalizing Constraints.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Feature selection (FS) is crucial for improving prediction accuracy and mitigating the curse of dimensionality in high-dimensional datasets.
• Neural networks excel at solving complex nonlinear learning problems.
• High-dimensional, low-sample-size data present significant challenges for effective feature selection.

Purpose of the Study:

• To propose a novel neural network-based feature selection (FS) approach.
• To develop a method that addresses the challenges of FS in high-dimensional, low-sample-size data.
• To enhance prediction accuracy and feature relevance identification.

Main Methods:

• A new neural network-based feature selection (FS) approach is proposed.
• The method incorporates two specific constraints to create a sparse FS layer.
• The approach is validated using extensive experiments on both synthetic and real-world datasets.

Main Results:

• The proposed sparse neural-network layer with normalizing constraints (SNeL-FS) method effectively selects important features.
• SNeL-FS demonstrates superior performance compared to conventional FS methods.
• The method shows particular efficacy on high-dimensional, low-sample-size data.

Conclusions:

• The SNeL-FS method offers a powerful new tool for feature selection in machine learning.
• This approach successfully handles the complexities of high-dimensional data.
• The findings suggest SNeL-FS can significantly improve model performance and interpretability.