Adaptive distributed support vector regression of massive data

Abstract

Massive datasets bring new challenges to traditional statistical inference, particularly in terms of memory restriction and computation time. Support vector regression is a robust and efficient estimation method. We first adopt smoothing techniques to develop smoothed support vector regression (S-SVR) estimation method. Then we propose distributed S-SVR (DS-SVR) algorithm for massive datasets. The proposed method solves the problems of memory restriction and computation time, and the resulting estimate can achieve the same efficiency as the estimator computed on all data. We also establish the asymptotic normality of the resulting estimate. In addition, we propose an adaptive learning process of parameters by using a combination of grid search and k − fold cross-validation, in which the optimal parameters ( λ , ϵ ) are automatically selected by each data. Finally, the performance of the proposed method is illustrated well by simulation studies.