Detecting influential observations by cluster analysis and Monte Carlo cross-validation

Abstract

The detection of influential observations is an essential step for building high performance models and has been recognized as an important and challenging task in many industrial and laboratorial applications. A new approach for detecting influential observations is developed based on their effect on partial least squares (PLS) modeling. In this method, we build a large number of PLS models by using Monte Carlo cross-validation (MCCV), and then perform principal component analysis (PCA) on the regression coefficients of these models. Because a model with influential observations is different from the one without influential observation, the series of PLS models cluster into different groups in principal component (PC) spaces, based on the different number of influential observations they contain. The influential observations can be therefore recognized according to the frequency number of each sample in each group. By three examples quantitatively modeling near-infrared (NIR) and Raman spectra, it was shown that the method can detect the influential observations intuitively and veraciously.