ASSESSING ROBUSTNESS OF CLASSIFICATION USING ANGULAR BREAKDOWN POINT.

Area of Science:

• Statistical Learning Theory
• Machine Learning Robustness

Background:

• Robustness is crucial for statistical techniques.
• Breakdown point is a standard robustness measure for regression.
• Existing breakdown point criteria are insufficient for classification problems.

Purpose of the Study:

• Propose a new criterion, the angular breakdown point, for quantifying classification robustness.
• Evaluate the robustness of binary large margin classification methods.
• Provide insights into the robustness of various classification techniques.

Main Methods:

• Introduced the novel angular breakdown point criterion.
• Analyzed robustness of binary large margin classification.
• Considered bounded and unbounded loss functions.
• Examined linear and kernel learning approaches.

Main Results:

• The angular breakdown point effectively quantifies classification robustness.
• Theoretical insights were gained into the robustness of large margin classifiers.
• Numerical results validated the theoretical findings.

Conclusions:

• The angular breakdown point is a valuable metric for assessing classification robustness.
• The study offers a deeper understanding of robustness in large margin classification.
• The proposed criterion can be applied to general classification methods.