Generalized SMOTE: A universal generation oversampling technique for all data types in imbalanced learning

Abstract

A common problem that arises when facing classification tasks is the class imbalance problem, which happens when one or more classes are heavily underrepresented compared to the rest, being usually those minority classes the ones of interest. A natural solution consists of correcting the imbalance by sampling methods, being Synthetic Minority Oversampling TEchnique (SMOTE) the most widely used method. In the same way as all other oversampling techniques, it relies on using distances/similarities in order to focus on the neighborhoods of minority samples in the synthetic samples generation procedure, thus it is meant for pure numerical data. Nevertheless, it is really common to collect categorical data or to discretize numeric attributes as a preprocessing step, being limited to random sampling approaches to correct imbalance. Some approaches have been proposed to deal with mixed-type data or pure categorical data, but they ignore part of the information of the samples or end up being almost random approaches. We propose GSMOTE, a generalization of SMOTE method, suitable for any data type. For the neighborhoods determination, the distance between samples is obtained by means of a trans formation of Gower’s General Similarity Coefficient into a novel General Distance Coefficient, in which the part corresponding to the way of measuring similarities between categories in categorical variables has been replaced by a recently presented similarity measure called Variable Entropy measure, inspired by Shannon’s Entropy. GSMOTE has been tested on six public imbalanced datasets, with different characteristics and imbalance levels.