Complex Localization in the Multiple Instance Learning Context

This paper introduces two approaches for solving Multiple Instance Problems (MIP) in which the traditional instance localization assumption is not met. We introduce a technique which transforms individual feature values in the attempt to align the data to the MIP localization assumption and a new MIP learning algorithm which identifies a region enclosing the majority (negative) class while excluding at least one instance from each positive (minority class) bag. The proposed methods are evaluated on synthetic datasets, as well as on a real-world manufacturing defect identification dataset. The real-world dataset poses additional challenges: data with noise, large imbalance and overlap.