Assessing robustness of classification using an angular breakdown point (Q1990584)

Classification's reliability is in the heart of the state of the art of machine learning methods. A novel measure of global reliability is proposed using population and sample angular breakdown points considering the difference between estimates with and without outliers. First, the authors consider linear learning with both bounded and unbounded loss functions, and the theoretical properties of an angular breakdown point for binary classification are studied. Next, an angular breakdown point for binary kernel learning (RKHS) with bounded or unbounded loss functions is considered. The efficiency of the proposed angular breakdown criterion is demonstrated on two datasets including a diagnostic breast cancer dataset.