Parameter-free classification in multi-class imbalanced data sets

Many applications deal with classification in multi-class imbalanced contexts. In such difficult situations, classical CBA-like approaches (Classification Based on Association rules) show their limits. Most CBA-like methods actually are One-Vs-All approaches (OVA), i.e., the selected classification rules are relevant for one class and irrelevant for the union of the other classes. In this paper, we point out recurrent problems encountered by OVA approaches applied to multi-class imbalanced data sets (e.g., improper bias towards majority classes, conflicting rules). That is why we propose a new One-Versus-Each (OVE) framework. In this framework, a rule has to be relevant for one class and irrelevant for every other class taken separately. Our approach, called fitcare, is empirically validated on various benchmark data sets and our theoretical findings are confirmed.