Metric Learning-Guided Least Squares Classifier Learning.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Multicategory classification problems require effective methods for distinguishing between classes.
• Discriminative Least Squares Regression (DLSR) uses an -dragging technique to improve classification margins.
• Existing methods may not fully leverage metric learning principles for error optimization.

Purpose of the Study:

• To propose a novel framework, the Metric Learning-guided Least Squares Classifier (MLG-LSC), for multicategory classification.
• To approach classification from a metric learning perspective, focusing on optimizing the error metric.
• To develop a classifier that implicitly performs error-dragging (e-dragging) for improved category separation.

Main Methods:

• Learning a unified metric matrix to minimize distances within categories and maximize distances between categories.
• Implementing an implicit error-dragging (e-dragging) mechanism within the least squares regression framework.
• Utilizing strictly convex optimization objectives to derive closed-form solutions for enhanced computational performance.

Main Results:

• The proposed MLG-LSC framework effectively learns a metric matrix for improved classification.
• The implicit e-dragging approach naturally captures relative category distance relationships.
• The method achieves superior computational performance due to closed-form solutions.

Conclusions:

• The MLG-LSC framework offers a valid and effective approach to multicategory classification.
• Metric learning provides a powerful perspective for optimizing least squares classifiers.
• The proposed method demonstrates competitive performance and computational advantages over existing techniques.