A Data Complexity Approach to Kernel Selection for Support Vector Machines