Abstract: In multi-instance learning, mechanism of making use of unlabeled instance would cut down training cost and increase generalization ability of the learner. Current algorithms perform semi-supervised multi-instance learning mainly by labeling each instance in bags and transferring multi-instance learning problem to singlcinstance semi supervised ones. In this paper we introduced a bag-level semi supervised learning framework with the idea that bag's label is determined by its instances and structure. With definition of multi-instance kernel, all bags(labeled and unlabeled) were used to calculate bag-level graph laplacian, which is a penalization term added to the optimization goal. We turned this problem into an optimization problem in RKHS and got a modified multi instance kernel function by unlabeled data as result that can be directly used in traditional kernel learning framework. We performed experiment in ALOI and Internet image datasets and compareed it with related algorithms. Experiment result shows that the proposed method can get the same accuracy as supervised counterpart with less labeled bags, and with the same labeled training data set the proposed method is of higher generalization ability.

Key words: Multi-instance learning, Semi-supervised learning, Multi-instance kernel, Support vector machine