PLS model building with missing data: New algorithms and a comparative study

[EN] New algorithms to deal with missing values in predictive modelling are presented in this article. Specifically, 2 trimmed scores regression adaptations are proposed, one from principal component analysis model building with missing data (MD) and other from partial least squares regression model...

ver descrição completa

Detalhes bibliográficos
Autores: Folch-Fortuny, Abel, Arteaga, Francisco, Ferrer, Alberto|||0000-0001-7244-5947
Formato: artículo
Fecha de publicación:2017
País:España
Recursos:Universitat Politècnica de València (UPV)
Repositorio:RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Idioma:inglés
OAI Identifier:oai:riunet.upv.es:10251/153468
Acesso em linha:https://riunet.upv.es/handle/10251/153468
Access Level:acceso abierto
Palavra-chave:Imputation
Missing data
Multivariate calibration
Partial least squares regression (PLS)
Trimmed scores regression (TSR)
ESTADISTICA E INVESTIGACION OPERATIVA
Descrição
Resumo:[EN] New algorithms to deal with missing values in predictive modelling are presented in this article. Specifically, 2 trimmed scores regression adaptations are proposed, one from principal component analysis model building with missing data (MD) and other from partial least squares regression model exploitation with missing values. Using these methods, practitioners can impute MD both in the explanatory/predictor and the dependent/response variables. Partial least squares is used here to build the multivariate calibration models; however, any regression method can be used after MD imputation. Four case studies, with different latent structures, are analysed here to compare the trimmed scores regression¿based methods against state-of-the-art approaches. The MATLAB code for these methods is also provided for its direct implementation at http://mseg.webs.upv.es, under a GNU license.