Relative Comparison Kernel Learning with Auxiliary Kernels

Finding a meaningful model of similarity between objects is vital for many machine learning tasks. In this work we consider the problem of learning similarity in the form of a positive semidefinite kernel matrix from relative comparisons. For many practical scenarios a large number of comparisons is hard to obtain, thus a complete model of similarity is difficult to learn. Previous attempts at solving this task only consider the relative comparisons in the learning process, but in almost all cases objects have other, easily extractable, information regarding their similarity to other objects. Our method uses these sources of auxiliary information in the learning of a kernel to supplement the few relative comparisons. We do this by learning a convex combination of kernels built from the auxiliary information and a unique kernel in which we learn the elements directly. Empirical results show that in the presence of few training relative comparisons our method learns kernels that can generalize better to unseen comparisons than methods that do not utilize auxiliary information.