刘一鹤，北京建筑大学

We propose a sequential method for implementing reconstruction regression. With a space-filling design of a small sample size, we use GP interpolation-based reconstruction regression to build initial prediction model. In the following steps, a new knot, corresponding to the maximum residual, is added, and the prediction model is updated at each iteration. This process stops when the generalized cross-validation criterion achieves a threshold. The proposed method has much less computational burden than standard GP-based and kernel methods, and is suitable for big data analysis. Numerical simulations are presented to illustrate our method.