Constructing Target Concept in Multiple Instance Learning Using Maximum Partial Entropy

摘要

Multiple instance learning, when instances are grouped into bags, concerns learning of a target concept from the bags without reference to their instances. In this paper, we advance the problem with a novel method based on computing the partial entropy involving only the positive bags using a partial probability scheme in the attribute sub-space. The evaluation highlights what could be obtained if information only from the positive bags is used, while the contributions from the negative bags are identified. The proposed method attempts to relax the dependency on the distribution of the whole probability of training data, but focus only on the selected subspace. Experimental evaluation explores the effectiveness of using maximum partial entropy in evaluating the merits between the positive and negative bags in the learning.