Efficiency versus Convergence of Boolean Kernels for On-Line Learning Algorithms