D3SC: Collaborative Research: Overcoming Challenges in Classification Near the Limit of Determination

点击翻译按钮获取中文摘要

With support from the Chemical Measurement & Imaging program in the Division of Chemistry, and partial co-funding from the Established Program to Stimulate Competitive Research (EPSCoR) and the Division of Mathematical Sciences, Professors Karl Booksh and Jocelyn Alcantara-Garcia at the University of Delaware, and Professor Barry Lavine at Oklahoma State University, are collaborating to improve the ability of hand-held chemical sensors for rapid sample classification. The problem is important, for example, for field analysis related to chemical forensics and art conservation, when the observed differences between two or more classes of interest are small compared to the natural variation among samples or among replicate measurements on a single sample. The team is developing advanced statistical and mathematical tools to enable quantitative determination of the statistical confidence with which the reliability of inferences can be assessed. The project entails combining information from two or more disparate sensors in order improve overall performance. Graduate and undergraduate students participating in this interdisciplinary research gain skills in advanced data analysis. These skills are in very high demand.This project is a collaborative effort aimed at investigating fundamental issues important to chemical modeling in modern measurement science: (1) improving classification model efficiency through variable selection, (2) assigning robust confidence levels to classifications that account for non-normal distributions of errors and class memberships, (3) increasing reliability of classification models when information from different sensors is available. The primary measurement tools are hand-held Laser Induced Breakdown Spectroscopy (LIBS) and X-Ray Fluorescence (XRF) data. This project probes the connections among variable selection, data fusion, optimization of instrumental parameters, and the performance of classification models. Targets include real-world classification applications where the class distribution and/or the measurement errors are not normally distributed. Nested bootstraps and genetic algorithms are being employed to solve this multilayered optimization problem. The developed methods will be modified as needed for PLS-DA, ANN-, and KNN-driven classifications. Resulting data sets will be made publicly available.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.