Minimum variance-embedded deep kernel regularized least squares method for one-class classification and its...

Area of Science:

• Machine Learning
• Deep Learning
• Biomedical Data Analysis

Background:

• Deep kernel learning is established for multi-class tasks, but less explored for one-class classification (OCC).
• OCC models require training data from a single class.
• Existing kernel regularized least squares (KRL) deep architectures for OCC exist.

Purpose of the Study:

• To introduce a novel deep kernel learning architecture for one-class classification (OCC).
• To enhance classifier generalization by embedding minimum variance information, reducing intra-class variance.
• To demonstrate efficacy on small datasets, including real-world biomedical applications.

Main Methods:

• Developed a novel deep kernel learning architecture by embedding minimum variance information into a KRL-based deep model.
• Evaluated the proposed classifier on 18 benchmark datasets (13 biomedical, 5 other).
• Compared performance against 16 state-of-the-art one-class classifiers.

Main Results:

• The proposed method effectively reduces intra-class variance, improving generalization.
• Demonstrated strong performance on small-size datasets, outperforming traditional deep learning methods.
• Achieved over 5% higher F1 score compared to state-of-the-art methods on biomedical benchmark datasets.
• Successfully applied to Alzheimer's disease detection and breast cancer detection from medical images.

Conclusions:

• The novel deep kernel learning approach with minimum variance embedding is effective for one-class classification.
• The method shows significant promise for biomedical applications with limited data availability.
• Outperforms existing state-of-the-art OCC methods, especially in the biomedical domain.