The gROC curve and the optimal classification.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Binary classification problems (BCP) involve assigning subjects to one of two groups based on a characteristic of interest.
• Traditional methods include binary regression and machine learning techniques like support vector machines.
• Receiver-operating characteristic (ROC) curves are standard tools for evaluating binary classifiers, assuming higher scores indicate higher probability of a positive outcome.

Purpose of the Study:

• To study, compare, and approximate transformations leading to efficient ROC (eROC) and generalized ROC (gROC) curves.
• To investigate the relationship between gROC and eROC curves.
• To propose non-parametric procedures for approximating gROC transformations.

Main Methods:

• Theoretical analysis to prove the equivalence of eROC and gROC curves under specific conditions (absence of relative maximum in optimal transformation).
• Investigation of gROC curve utility in theoretical models.
• Development and application of two non-parametric procedures for approximating gROC transformations.

Main Results:

• Demonstrated that eROC and gROC curves are equivalent when the optimal transformation lacks a relative maximum.
• Explored the theoretical relationship between gROC and eROC curves.
• Proposed and evaluated two non-parametric approximation methods for gROC transformations using Monte Carlo simulations and real-data analysis.

Conclusions:

• The study establishes conditions for the equivalence of eROC and gROC curves, offering a unified perspective.
• The proposed non-parametric methods provide practical tools for approximating gROC transformations.
• The findings are illustrated with simulations and real-world data, highlighting the utility of gROC curves in binary classification.