Robust BMPM training based on second-order cone programming and its application in medical diagnosis

Area of Science:

• Machine Learning
• Optimization
• Medical Data Analysis

Background:

• Imbalanced learning tasks pose challenges for traditional classifiers.
• The Biased Minimax Probability Machine (BMPM) offers a worst-case bound on misclassification probability.
• Existing BMPM methods may have limitations in accuracy and robustness.

Purpose of the Study:

• To develop a novel extension training algorithm for BMPM using Second-Order Cone Programming (SOCP).
• To enhance the accuracy and robustness of BMPM by relaxing original assumptions.
• To apply the improved BMPM to medical diagnosis for better sensitivity.

Main Methods:

• Reformulated the biased classification model into an SOCP problem for efficient solving with global optima guarantee.
• Developed a novel SOCP-based training algorithm for BMPM (BMPMSOCP).
• Evaluated BMPMSOCP on medical diagnosis tasks, focusing on improving sensitivity.

Main Results:

• The proposed BMPMSOCP scheme demonstrates superior effectiveness and robustness compared to traditional methods and the original BMPM.
• Empirical results confirm the method's ability to handle imbalanced classification problems in medical contexts.
• The SOCP approach ensures efficient computation and guarantees global optima.

Conclusions:

• The SOCP-based BMPM is a more accurate and robust approach for imbalanced classification, particularly in medical diagnosis.
• This method effectively improves sensitivity for critical classes, outperforming previous BMPM algorithms.
• The study validates the practical utility of BMPMSOCP in real-world medical applications.