Salient instance selection for multiple-instance learning

Abstract

Multiple-instance learning (MIL) is a variant of traditional supervised learning, where training examples are bags of instances. In this learning framework, only the labels of bags are known while the labels of instances in bags are unknown. This ambiguity in labels of instances leads to significant challenges in MIL. In this paper, we propose an efficient instance selection method to solve this problem, called Salient Instance Selection for Multiple-Instance Learning (MILSIS). MILSIS has two roles: first, selecting discriminative instances and eliminating redundant or irrelevant instances from each bag; second, selecting an instance prototype from each positive bag to construct an embedding space in order to convert the MIL problem to the standard single instance learning problem. Accordingly, based on the first role, we present two novel MIL methods, called MILSIS-kNN-C and MILSIS-kNN-B; based on the second role, we present another new MIL method, called MILSIS-SVM. Experimental results on some synthetic and benchmark data-sets demonstrate the effectiveness of our methods as compared to others. © 2012 Springer-Verlag.