Cost-sensitive one-vs-one ensemble for multi-class imbalanced data

Learning from imbalanced data poses significant challenges for machine learning algorithms, as they need to deal with uneven distribution of examples in the training set. As standard classifiers will be biased towards the majority class there exist a need for specific methods than can overcome this single-class dominance. Most of works concentrated on binary problems, where majority and minority class can be distinguished. But a more challenging problem arises when imbalance is present within multi-class datasets, as relations between classes tend to complicate. One class can be a minority class for some, while a majority for others. In this paper, we propose an efficient method for handling such scenarios, that combines the problem decomposition with cost-sensitive learning. According to divide-and-conquer rule, we decompose our multi-class data into a number of binary subproblems using one-versus-one approach. To each simplified task we delegate a cost-sensitive neural network with moving threshold. It relies on scaling the output of the classifier with a given cost function. This way, we adjust our support functions towards the minority class. We propose a novel method for automatically determining the cost, based on the Receiver Operating Characteristic (ROC) curve analysis. This way we can estimate the cost matrix for each class pair independently. Then using a dedicated classifier fusion approach, we reconstruct the original multi-class problem. Experimental analysis backed-up with statistical testing clearly proves that such an approach is superior to state-of-the art ad-hoc and decomposition methods used in the literature.