A Geometric Algorithm for Scalable Multiple Kernel Learning

We present a geometric formulation of the Mul-tiple Kernel Learning (MKL) problem. To do so, we reinterpret the problem of learning ker-nel weights as searching for a kernel that max-imizes the minimum (kernel) distance between two convex polytopes. This interpretation com-bined with novel structural insights from our geo-metric formulation allows us to reduce the MKL problem to a simple optimization routine that yields provable convergence as well as quality guarantees. As a result our method scales effi-ciently to much larger data sets than most prior methods can handle. Empirical evaluation on eleven datasets shows that we are significantly faster and even compare favorably with a uni-form unweighted combination of kernels.