Semi-Supervised Discriminative Classification Robust to Sample-Outliers and Feature-Noises.

Area of Science:

• Machine Learning
• Medical Image Analysis
• Data Science

Background:

• Discriminative models struggle with real-world data containing sample-outliers and feature-noises.
• Robust methods for simultaneous outlier and noise detection are underrepresented in current literature.
• Semi-supervised learning can enhance denoising by leveraging both labeled and unlabeled data to better model data geometry.

Purpose of the Study:

• To propose a semi-supervised robust discriminative classification method.
• To simultaneously detect and address sample-outliers and feature-noises.
• To improve classification performance in challenging real-world datasets, including medical imaging.

Main Methods:

• Utilized a least-squares formulation of linear discriminant analysis (LDA).
• Developed a semi-supervised approach incorporating both labeled training and unlabeled testing data.
• Tested the method on synthetic, benchmark semi-supervised, and neurodegenerative disease (Parkinson's, Alzheimer's) datasets.

Main Results:

• The proposed method demonstrated superior performance compared to baseline and state-of-the-art approaches.
• Achieved higher accuracy and Area Under the ROC Curve (AUC) on tested datasets.
• Effectively handled datasets with sample-outliers and feature-noises, particularly neuroimaging data.

Conclusions:

• The novel semi-supervised robust method effectively addresses outliers and noise in discriminative classification.
• The approach shows significant promise for applications in medical image analysis and disease diagnosis.
• Leveraging unlabeled data alongside robust techniques enhances model generalization in the presence of data imperfections.