Graph convolutional network-based feature selection for high-dimensional and low-sample size data.

Area of Science:

• Machine Learning
• Data Science
• Bioinformatics

Background:

• Feature selection is crucial for dimension reduction in machine learning.
• Traditional methods struggle with high-dimensional and low-sample size (HDLSS) data due to overfitting.
• Overfitting leads to poor model generalization and unreliable results.

Purpose of the Study:

• To introduce a novel deep learning-based feature selection method for HDLSS data.
• To address the limitations of existing feature selection techniques in challenging data settings.
• To improve the accuracy and robustness of models built on HDLSS datasets.

Main Methods:

• Developed GRaph Convolutional nEtwork feature Selector (GRACES), a deep learning approach.
• GRACES leverages graph convolutional networks to exploit latent sample relationships.
• Employs overfitting-reducing techniques for robust feature identification.

Main Results:

• GRACES significantly outperforms existing feature selection methods.
• Demonstrated superior performance on both synthetic and real-world HDLSS datasets.
• Achieved substantial decreases in optimization loss through iterative feature selection.

Conclusions:

• GRACES offers a powerful solution for feature selection in HDLSS data.
• The method effectively mitigates overfitting challenges inherent in such datasets.
• Publicly available code facilitates reproducibility and adoption.