蒙特卡洛交叉验证用于偏最小二乘建模数据质量的评价（英文）

Evaluation of calibration data for partial least squares modeling by using Monte Carlo cross validation

基于蒙特卡洛交叉验证(MCCV)建立了一种用于近红外光谱偏最小二乘建模数据质量的评价方法。该方法首先通过蒙特卡洛交叉验证计算交叉验证均方根误差(RMSECV),同时计算交叉验证中建模样本的预测误差,记为建模样本的均方根误差(RMSECVc)。如果数据中部存在奇异样本、噪声、非线性相应等干扰因素,RMSECV和RMSECVc随因子数的变化应该保持一致,否则,二者的变化趋势将不同。因此,利用RMSECV和RMSECVc随因子数的变化趋势即可对数据的质量进行评价。采用模拟数据和12组实际样品的数据对该方法进行了考察,并对四组实际数据中的奇异样本进行分析,说明了方法的效果。本文为偏最小二乘建模方法提供了一种数据质量的评价方法。

A method based on Monte Carlo cross validation （MCCV） is proposed for evaluation of calibration data for partial least squares （PLS） regression. In the method, the root mean squared error of cross validation （RMSECV） is calculated as usual using the prediction errors in the MCCV, and another RMSECV is calculated using the prediction errors of the samples that are selected for building the models. The latter is denoted as RMSECVc. If there is no interfering factor in the calibration data, e.g., outlier, noise, or nonlinear responses, the variation of RMSECV and RMSECVc with the latent variable （LV） number will be in a same trend. Otherwise, there will be a difference between the two values after an LV number when the interfering factors are encoded in the model. Therefore, a comparison of the RMSECV and RMSECVc curves can be used for detecting the interfering factors contained in the calibration data. A simulated dataset and 12 real near infrared spectroscopic datasets were used to test the proposed method. The effect of outliers in four real datasets was analyzed. The results show that the method provides a useful tool for evaluation of the calibration dataset and the quality of PLS models.