Marginalized Multi-Instance Kernels

Support vector machines (SVM) have been highly successful in many machine learning problems. Recently, it is also used for multi-instance (MI) learning by employing a kernel that is defined directly on the bags. As only the bags (but not the instances) have known labels, this MI kernel implicitly assumes all instances in the bag to be equally important. However, a fundamental property of MI learning is that not all instances in a positive bag necessarily belong to the positive class, and thus different instances in the same bag should have different contributions to the kernel. In this paper, we address this instance label ambiguity by using the method of marginalized kernels. It first assumes that all the instance labels are available and defines a label-dependent kernel on the instances. By integrating out the unknown instance labels, a marginalized kernel defined on the bags can then be obtained. A desirable property is that this kernel weights the instance pairs by the consistencies of their probabilistic instance labels. Experiments on both classification and regression data sets show that this marginalized MI kernel, when used in a standard SVM, performs consistently better than the original MI kernel. It also outperforms a number of traditional MI learning methods.