Deformation of log-likelihood loss function for multiclass boosting

Area of Science:

• Machine Learning
• Statistical Learning Theory

Background:

• Classification problems rely on minimizing empirical loss functions to estimate decision functions.
• Existing loss functions may not adequately address issues like mislabeling or outliers in multiclass settings.

Purpose of the Study:

• To propose and investigate a novel class of deformed log-likelihood loss functions for multiclass classification.
• To develop a robust loss function against outliers and a statistical model for mislabeling.
• To explore the mathematical properties and computational advantages of these new loss functions.

Main Methods:

• Deformation of the standard log-likelihood loss function.
• Development of a boosting algorithm utilizing a pseudo-loss for minimization.
• Mathematical characterization of the proposed loss functions.
• Extension of existing robust loss functions for binary classification to multiclass settings.

Main Results:

• The proposed deformed log-likelihood loss functions establish a clear link between decision functions and conditional probabilities.
• These functions ensure statistical consistency of the classification error rate.
• A novel model of mislabeling is presented, applicable to medical diagnostics.
• A robust loss function effective against outliers in multiclass classification is introduced.

Conclusions:

• The deformed log-likelihood loss functions offer significant advantages in multiclass classification, including improved statistical properties and interpretability.
• The developed robust loss function and mislabeling model provide effective solutions for handling noisy data.
• The findings contribute to advancing the theory and practice of robust and reliable classification methods.