A Robust AUC Maximization Framework With Simultaneous Outlier Detection and Feature Selection for Positive-Unlabeled...

Area of Science:

• Machine Learning
• Data Science
• Bioinformatics

Background:

• Positive-unlabeled (PU) classification is prevalent in healthcare, text analysis, and bioinformatics.
• PU classification faces challenges with imbalanced data, mislabeled samples, and corrupted features.
• Existing methods struggle with complex datasets common in real-world PU scenarios.

Purpose of the Study:

• To develop a robust learning framework for PU classification.
• To address data complexity, mislabeled samples, and corrupted features in PU learning.
• To provide theoretical insights and practical guidance for PU model training.

Main Methods:

• Proposed a unified framework integrating area under the curve (AUC) maximization, outlier detection, and feature selection.
• Developed theoretical generalization error bounds for the proposed PU learning model.
• Validated the framework through empirical comparisons and real-world case studies.

Main Results:

• The unified framework effectively handles biased labels, excludes mislabeled samples, and removes corrupted features.
• Theoretical analysis provides insights into model performance and optimal sample size for training.
• Demonstrated significant effectiveness in surgical site infection (SSI) and EEG seizure detection tasks.

Conclusions:

• The proposed robust learning framework enhances PU classification performance, particularly for complex and noisy datasets.
• Theoretical bounds offer practical guidance for effective model training and data selection.
• The method shows strong potential for real-world applications in healthcare and bioinformatics.