Polichotomies on Imbalanced Domains by One-per-Class Compensated Reconstruction Rule

A key issue in machine learning is the ability to cope with recognition problems where one or more classes are under-represented with respect to the others. Indeed, traditional algorithms fail under class imbalanced distribution resulting in low predictive accuracy over the minority classes. While large literature exists on binary imbalanced tasks, few researches exist for multiclass learning. In this respect, we present here a new method for imbalanced multiclass learning within the One-per-Class decomposition framework. Once the multiclass task is divided into several binary tasks, the proposed reconstruction rule discriminates between safe and dangerous classifications. Then, it sets the multiclass label using information on both data distributions and classification reliabilities provided by each binary classifier, lowering the effects of class skew and improving the performance. We favorably compare the proposed reconstruction rule with the standard One-per-Class method on ten datasets using four classifiers.