On the Robustness of Feature Selection with Absent and Non-observed Features

To improve upon early detection of Classical Swine Fever, we are learning selective Naive Bayesian classifiers from data that were collected during an outbreak of the disease in the Netherlands. The available dataset exhibits a lack of distinction between absence of a clinical symptom and the symptom not having been addressed or observed. Such a lack of distinction is not uncommonly found in biomedical datasets. In this paper, we study the effect that not distinguishing between absent and non-observed features may have on the subset of features that is selected upon learning a selective classifier. We show that while the results from the filter approach to feature selection are quite robust, the results from the wrapper approach are not.