The efficiency of support vector machine in practice is closely related to the optimal selection of kernel functions and their hyper-parameters. A novel kernel, namely the arctangent kernel, is proposed in this paper. Compared with the Gaussian kernel, the new proposed kernel has a quick similarity descent in the neighborhood of the inspection sample and a moderate similarity descent toward the infinity of the inspection sample. The experimental results on two simulated data sets and some UCI data sets show that the new proposed kernel function has better effectiveness and robustness compared with the polynomial kernel, the Gaussian kernel, the exponential radial basis function, and the former proposed kernel with moderate decreasing.