M-ary Rank Classifier Combination: A Binary Linear Programming Problem

Area of Science:

• Machine Learning
• Computer Vision
• Medical Diagnostics

Background:

• Classifier combination aims to improve predictive performance by integrating multiple classifiers.
• Rank-based outputs offer richer information than single class predictions, especially in multi-class problems.

Purpose of the Study:

• To propose a novel method for combining weak rank classifiers.
• To formulate classifier decision combination as a discrete optimization problem.
• To develop a consensus measure for assessing classifier agreement.

Main Methods:

• Formulating classifier combination as a discrete optimization problem using candidate class rankings.
• Developing a consensus function that maps individual rankings to a combined ranking.
• Proposing an information measure to quantify classifier agreement without requiring training data.

Main Results:

• The proposed approach transforms the combination problem into a binary linear programming problem.
• A global optimum solution is achievable for the consensus ranking.
• Classification rates are significantly improved by combining rank classifiers.

Conclusions:

• Combining weak rank classifiers offers a robust strategy for enhancing classification performance.
• The developed method is effective and easily implementable, demonstrated on cervical cancer detection data.
• The consensus measure provides a valuable tool for evaluating the reliability of the combination rule.