Addressing imbalanced classification with instance generation techniques: IPADE-ID

Abstract

A wide number of real word applications presents a class distribution where examples belonging to one class heavily outnumber the examples in the other class. This is an arduous situation where standard classification techniques usually decrease their performance, creating a handicap to correctly identify the minority class, which is precisely the case under consideration in these applications.In this work, we propose the usage of the Iterative Instance Adjustment for Imbalanced Domains (IPADE-ID) algorithm. It is an evolutionary framework, which uses an instance generation technique, designed to face the existing imbalance modifying the original training set. The method, iteratively learns the appropriate number of examples that represent the classes and their particular positioning. The learning process contains three key operations in its design: a customized initialization procedure, an evolutionary optimization of the positioning of the examples and a selection of the most representative examples for each class.An experimental analysis is carried out with a wide range of highly imbalanced datasets over the proposal and recognized solutions to the problem. The results obtained, which have been contrasted through non-parametric statistical tests, show that our proposal outperforms previously proposed methods.