Optimizing area under the ROC curve using semi-supervised learning.

Area of Science:

• Machine Learning
• Computer Science
• Medical Imaging Analysis

Background:

• Receiver Operating Characteristic (ROC) analysis is crucial for evaluating binary classification systems.
• Area Under the Curve (AUC) is a key metric for assessing classifier performance.
• Traditional AUC optimization relies solely on labeled data, limiting its effectiveness.

Purpose of the Study:

• To propose novel semi-supervised learning receiver operating characteristic (SSLROC) algorithms for AUC optimization.
• To enhance classifier training by incorporating unlabeled test samples.
• To improve the adaptability of decision boundaries to data distributions.

Main Methods:

• Developed two SSLROC algorithms (SSLROC1 and SSLROC2) inspired by semi-supervised and transductive learning.
• Formulated the AUC optimization as a semi-definite programming problem using margin maximization.
• Incorporated unlabeled samples as constraints based on their ranking relationships with labeled data.

Main Results:

• SSLROC algorithms demonstrated significant improvements over state-of-the-art methods on 34 diverse datasets.
• The methods effectively adapt decision boundaries to both labeled training and unlabeled test data distributions.
• Promising results were observed when applied to a CT colonography dataset for colonic polyp classification.

Conclusions:

• The proposed SSLROC algorithms offer a powerful approach to AUC optimization by effectively utilizing unlabeled data.
• These methods enhance classifier performance by considering the distribution of both labeled and unlabeled samples.
• SSLROC algorithms show potential for improving diagnostic accuracy in medical applications like polyp detection.