Improving representation of the positive class in imbalanced multiple–instance learning

Abstract

In standard supervised learning, the problem of learning from imbalanced data has been addressed to improve the performance of learning algorithms in the presence of underrepresented data. However, in Multiple-Instance Learning (MIL), where the imbalance problem is more complex, there is little discussion about it. Motivated by the need of further studies, we discuss the multiple-instance imbalance problem and propose a method to improve the representation of the positive class. Our approach looks for the target concept in positive bags and tries to strength it using an oversampling technique while removes the borderline (ambiguous) instances in positive and negative bags. We evaluate our method on several standard MIL benchmarking data sets in order to show its ability to get an enhanced representation of the positive class.