Abstract: The successful key, the prediction of displacement time series of landslide based on support vector machines (SVM), is designing or choosing a appropriate kernel function. The characters of classical displacement time series of landslide are analyzed; and some general properties of Mercer kernels are discussed. And we use a number of closure properties to create optimal kernels for the SVM prediction of landslide displacement data. The efficiencies of designed kernels are tested by predicting three groups of displacement time series. All experimental results show LPG and MPG kernels outperform the standard kernels, and the LPG kernel is of efficient about the regression and prediction of complex displacement time series. The MPG kernel is appropriate for simple deformation datasets with obvious laws. Furthermore, some good experiences are obtained about kernel parameters choice by investigating interconnection between different parameters and models learn performance.

Key words: kernel function, feature space, prediction, displacement time series, landslide