Using random forest for reliable classification and cost-sensitive learning for medical diagnosis

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Machine learning classifiers often lack reliability indicators for predictions.
• This limitation hinders their use in critical fields like medical diagnosis where errors have severe consequences.
• Standard classifiers assume equal misclassification costs, which is often unrealistic in medical settings.

Purpose of the Study:

• To develop a reliable machine learning classifier for critical applications.
• To address the challenge of varying misclassification costs in sensitive domains.
• To enhance prediction certainty and control error rates.

Main Methods:

• Integration of a modified random forest classifier with the conformal predictor scheme.
• Utilizing Kolmogorov complexity within the conformal predictor for sample randomness assessment.
• Development of a label-conditional predictor to handle cost-sensitive classification and class-specific confidence levels.

Main Results:

• The proposed method demonstrates a well-calibrated property, aligning accuracy with predefined confidence levels.
• The label-conditional extension effectively manages differential misclassification costs across classes.
• Experiments confirm the classifier's ability to control specific risks for different classes in benchmark and real-world datasets.

Conclusions:

• Employing a random forest outlier measure as a nonconformity measure enhances predictor performance.
• The label-conditional classifier offers a novel solution for cost-sensitive learning by setting label-wise confidence levels.
• Minimizing misclassification risk is achieved through tailored confidence levels for each class.