Kernel Selection in Support Vector Machines Using Gram-Matrix Properties

We describe an approach to kernel selection in Support Vector Machines (SVMs) driven by the Gram matrix. Our study extracts properties from this matrix (e.g., Fisher’s discriminant, Bregman’s divergence) using different kernel functions (linear, polynomial, Gaussian, Laplacian, Bessel and ANOVARBF), and incorporates such properties as meta-features within a meta-learning framework. The goal is to predict the best kernel in SVMs. Results show how introducing a new metafeature, Distance Ratio, capturing inter-class and intra-class distances in the feature space, yields substantial improvements during kernel selection.