Groupwise Retargeted Least-Squares Regression.

Area of Science:

• Machine Learning
• Computer Science
• Pattern Recognition

Background:

• Multicategory classification is a fundamental task in machine learning.
• Existing methods like Retargeted Least Squares Regression (ReLSR) have limitations in regularization.
• There is a need for improved classification models with enhanced discriminative power.

Purpose of the Study:

• To propose a novel groupwise retargeted least squares regression (GReLSR) model.
• To enhance multicategory classification by introducing class-specific regularization.
• To improve upon the performance of existing ReLSR and discriminative least-squares regression methods.

Main Methods:

• Developed a new formulation of ReLSR explicitly expressing translation values.
• Introduced a groupwise constraint to the ReLSR framework, creating the GReLSR model.
• Incorporated additional regularization to restrict translation values within the same class.

Main Results:

• The proposed GReLSR model demonstrated superior performance on various machine learning datasets.
• GReLSR effectively utilizes regularization to ensure similarity of translation values within classes.
• Discriminative least-squares regression is shown to be a special case of ReLSR within the new formulation.

Conclusions:

• The GReLSR model offers a significant advancement in multicategory classification.
• The groupwise constraint effectively enhances the discriminative capability of least-squares regression.
• GReLSR represents a new state-of-the-art approach for multicategory classification tasks.