Research on Screening and Evaluating Virtual Samples by Multi-Gaussian Process Regression Models

On the issue of virtual sample generation, researchers are now generally focusing on how to generate more virtual samples, while neglecting the evaluation and screening of virtual samples. In this work, a virtual sample evaluation and screening method based on multiple Gaussian process regression (GPR) is proposed. In this method, we firstly screen the virtual samples for outliers according to partial priori knowledge to get the candidate samples. Then the candidate samples are evaluated one by one according to the prediction results of multiple GPR models. Samples whose outputs were within the confidence range of the predicted outputs of one GPR model were regarded as passing the test of that model. In the screening stage, the candidate samples were screened according to the number of times they passed the tests and the actual modeling needs. By this way, we can get optimized virtual samples which can maintain or even improve the accuracy of prediction model with minimum number of virtual samples. Finally, the validity of this virtual sample screening method was verified on the dataset of extracting liquiritin by dual-frequency ultrasound.