Improved Loss Bounds For Multiple Kernel Learning

We propose two new generalization error bounds for multiple kernel learning (MKL). First, using the bound of Srebro and Ben-David (2006) as a starting point, we derive a new version which uses a simple counting argument for the choice of kernels in order to generate a tighter bound when 1-norm regularization (sparsity) is imposed in the kernel learning problem. The second bound is a Rademacher complexity bound which is additive in the (logarithmic) kernel complexity and margin term. This dependency is superior to all previously published Rademacher bounds for learning a convex combination of kernels, including the recent bound of Cortes et al. (2010), which exhibits a multiplicative interaction. We illustrate the tightness of our bounds with simulations.