Data-dependent and Scale-Invariant Kernel for Support Vector Machine Classification

Kernel similarity function allows a Support Vector Machine (SVM) classifier to learn the maximum margin hyperplane in a higher dimensional space where two classes are linearly separable without explicitly mapping the data. Most existing kernel functions (e.g., RBF) use spatial positions of two data instances in the input space to compute their similarity. These kernels are data distribution independent and sensitive to data representation (i.e., units/scales used to measure/express data). Since this can be unknown in many real-world applications, a careful selection of a suitable kernel is required for a given problem. In this paper, we present a new kernel function based on probability data mass that is both data-dependent and scale-invariant. Our empirical results show that the proposed SVM kernel outperforms popular existing kernels.