Wavelet analysis based sparse LS-SVR for time series data

Due to the performances of low computational cost and excellent generalization capability, Least squares support vector regression (LS-SVR) has been successfully applied to function estimation and forecasting problems. However, in comparison to SVR, LS-SVR loses the sparseness and has worse robustness for large training samples. In this paper, a sparse LS-SVR is proposed for the regression of large time series data. The signal features are extracted by using the multi-scale decomposition and wavelet denoising for training sample set. Based on the reconstructed signal, the importance of training samples is determined and the sparseness is imposed to LS-SVR. The typical benchmark functions are employed to evaluate our proposed algorithm. The experimental results show this algorithm can not only reduce the number of training samples significantly, but also eliminate noise interference.