Multiple Instance Transfer Learning

Transfer Learning is a very important branch in both Machine Learning and Data Mining. Its main objective is to transfer knowledge across domains, tasks and distributions that are similar but not the same. Currently, almost all of the transfer learning methods are designed to deal with the traditional single instance learning problems. However, in many real-world applications, such as drug design, Localized Content Based Image Retrieval (LCBIR), Text Categorization, we have to deal with multiple instance problems, where training patterns are given as {\em bags} and each bag consists of some \emph{instances}. This paper formulates a novel Multiple Instance Transfer Learning (MITL) problem and suggests a method to solve it. An extensive set of empirical results demonstrate the advantages of the proposed method against several existed ones.