A Hybrid Sampling Method for Imbalanced Data

With the diversification of applications and the emergence of new trends in challenging applications such as in the computer vision domain, classical machine learning systems usually perform poorly while confronting two common problems: the training data of negative examples, which outnumber the positive ones, and the large intra-class variations. These problems lead to a drop in the system performances. In this work, we propose to improve the classification accuracy in the case of imbalanced training data by equally balancing a training data set using a hybrid approach which consists in over-sampling the minority class using a "SMOTE star topology", and under-sampling the majority class by removing instances that are considered less relevant. The feature vector deletion has been performed with respect to intra-class variations, based on the distribution criterion. The experimental results, achieved in biometric data, show that the proposed approach significantly improves the overall performances measured in terms of true-positive rate.