Local Regression by Forming Matrix-Matched Calibration Sets Using Virtual Reality Sample Selection

Linear regression for quantitative analysis commonly uses a calibration model based on a large global sample set spanning a substantial domain of calibration sample variance. Because of the large variance, determining accurate analyte amounts present in new target prediction samples is challenging. Instead, calibration samples closely matrix-matched to each target sample can be selected to generate localized linear regression models for accurate target sample analyte predictions. The difficulty with local regression in analytical chemistry is how to select calibration samples that bracket each target sample over small respective ranges of innate matrix-matched measured responses and analyte amounts. Recently developed was a robust autonomous local sample selection algorithm called local adaptive fusion regression (LAFR) that can realize the two-goaled sample selection. The method is based on the physicochemical responsive integrated similarity measure (PRISM) algorithm to compute hundreds of similarity measures between a target sample and a global library sample set. Depending on the library size, LAFR can require extensive time-consuming calculations and this paper compares LAFR to an easier more intuitive computer-human hybrid technique using virtual reality (VR) for local model sample selection. Recently explored with VR were successful applications to various chemometric data analysis scenarios relying on the user's lifetime problem-solving training and inherent pattern recognition skills. Like LAFR, these studies and this study use PRISM to compute sample similarity measures. These similarity values are further rendered into VR sample glyph features providing intuitive visualization of sample matrix effects, ie, based on glyph appearances, the user selects matrix-matched calibration samples in VR localized to each target sample. Presented are partial least squares (PLS) prediction results from LAFR and VR selected calibration sets localized to fifteen target samples across three near infrared (NIR) data sets. By using VR, prediction errors are generally the same or better than LAFR without the lengthy LAFR calculations. Also presented is a discussion on the Rashomon effect supporting the substantial differences between the LAFR and VR regression vector shapes and magnitudes but the models predict similarly. Included in the discussion is the inability to interpret the models.