Clustering-based multiple instance learning with multi-view feature

Multi-instance learning (MIL) is a special kind of classifi cation problem where samples (called \instances") are grouped into bags and labels are given only on bag level instead of instance level. From the expert and intelligent system perspective, MIL does not require full ground-truth labels which helps to reduce the cost of data labeling in real tasks. Many inviting problems such as image classification, video annotation and object detection can be formulated in MIL frameworks. In this paper, we propose a similarity-based method with clustering in a multi-view feature manner to solve MIL problems efficiently. The clustering is introduced as a novel strategy for instance selection. Considering unlabeled data, clustering methods t well for obtaining the hidden structure information in feature space, and thus we utilize a clustering-based strategy to exploit such information on discovering positive instances. To fully use the original input data, we further develop our method in multi-view feature space, in order to make our model capture information from different feature space and provide support for the whole system. Experiments on benchmark datasets for classfi cation have been conducted, and we also perform image retrieval to examine the extended multi-view model on a large image dataset, MIL-NUS-WIDE. The promising results demonstrate the effectiveness of our clustering-based MIL (CMIL) model.