Multi-class Imbalanced Data-Sets with Linguistic Fuzzy Rule Based Classification Systems Based on Pairwise Learning

In a classification task, the imbalance class problem is present when the data-set has a very different distribution of examples among their classes. The main handicap of this type of problem is that standard learning algorithms consider a balanced training set and this supposes a bias towards the majority classes. In order to provide a correct identification of the different classes of the problem, we propose a methodology based on two steps: first we will use the one-vs-one binarization technique for decomposing the original data-set into binary classification problems. Then, whenever each one of these binary subproblems is imbalanced, we will apply an oversampling step, using the SMOTE algorithm, in order to rebalance the data before the pairwise learning process. For our experimental study we take as basis algorithm a linguistic Fuzzy Rule Based Classification System, and we aim to show not only the improvement in performance achieved with our methodology against the basic approach, but also to show the good synergy of the pairwise learning proposal with the selected oversampling technique.