Evidential Hybrid Re-sampling for Multi-class Imbalanced Data

International audienceLearning from class-imbalanced datasets has gained substantial attention in the machine learning community, leading to solutions for healthcare, security, banking, etc. Specifically, binary imbalanced problems has received the most interest in the field. Yet, there has been little emphasis given to dealing with multi-class imbalance learning. Data imbalance can significantly worsen the classification performance, especially in the presence of other data difficulties such as uncertainty, i.e., ambiguous samples and noise. In this paper, we present an evidential hybrid re-sampling method for dealing with class imbalance in the multi-class setting. This technique uses the evidence theory to assign a soft label to each object. This evidential modeling provides more information about each object's region, which improves the selection of objects in both undersampling and oversampling. Our approach firstly selects ambiguous majority instances for undersampling, then oversamples minority objects through the generation of synthetic examples in borderline regions to better improve minority class borders. An adjustment has also been integrated in order to avoid excessive oversampling and undersampling. Benchmarking results have shown significant improvement of G-Mean of AUC metrics over other popular re-sampling methods