Bayesian multiple imputation approaches for one-class classification

Abstract

One-Class Classifiers build classification models in the absence of negative examples, which makes it harder to estimate the class boundary. The predictive accuracy of one-class classifiers can be exacerbated by the presence of missing data in the positive class. In this paper, we propose two approaches based on Bayesian Multiple Imputation (BMI) for imputing missing data in the one-class classification framework called Averaged BMI and Ensemble BMI. We test and compare our approaches against the common method of Mean imputation and Expectation Maximization on several datasets. Our preliminary experiments suggest that as the missingness in the data increases, our proposed imputation approaches can do better on some data sets. © 2012 Springer-Verlag.