Constructing support vector machines with missing data

Support vector machine classification (SVM) is a statistical learning method which easily accommodates large numbers of predictors and can discover both linear and non-linear relationships between the predictors and outcomes. A common challenge is constructing an SVM when the training set includes observations with missing predictor values. In this paper, we identify when missing data can bias an SVM classifier. Because the missing data mechanisms which bias SVMs differ from the traditional framework of missing-at-random and missing-not-at-random, we argue for an SVM specific framework for understanding missing data. Further, we compare a number of missing data strategies for SVMs in a simulation study and real data example, and we make recommendations for SVM users based on the simulation study. ∗Department of Biostatistics, Vanderbilt University School of Medicine †Department of Biostatistics, University of North Carolina at Chapel Hill ‡Public Health Sciences Division, Fred Hutchinson Cancer Research Center