An embedded feature selection method based on generalized classifier neural network for cancer classification.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Gene selection is crucial for high-dimensional microarray data classification.
• Existing linear sparse learning models struggle with non-linear gene interactions and multi-class problems.
• Current methods often create overlapping groups, increasing data dimensionality.

Purpose of the Study:

• To propose a novel neural network-based embedded feature selection method for gene expression data.
• To address the limitations of linear models in capturing non-linear gene interactions.
• To develop an explainable and effective feature selection approach for both binary and multi-class problems.

Main Methods:

• A Generalized Classifier Neural Network (GCNN) was adopted for its explainable architecture.
• A Weighted GCNN (WGCNN) approach was developed, integrating feature weighting into neural network training.
• Statistical guided dropout was implemented at the input layer to prevent overfitting in high-dimensional data.

Main Results:

• WGCNN demonstrated effectiveness in representing non-linear relationships within gene expression data.
• The method showed strong performance across seven microarray datasets for binary and multi-class classification.
• Experimental validation confirmed WGCNN's superiority in F1 score and feature selection efficiency compared to state-of-the-art methods.

Conclusions:

• WGCNN offers an effective and explainable solution for gene selection in high-dimensional microarray data.
• The proposed method overcomes limitations of linear models by capturing non-linear interactions.
• WGCNN provides a robust approach for both binary and multi-class classification tasks in bioinformatics.