Selective Classifiers Can Be Too Restrictive: A Case-Study in Oesophageal Cancer

Real-life datasets in biomedicine often include missing values. When learning a Bayesian network classifier from such a dataset, the missing values are typically filled in by means of an imputation method to arrive at a complete dataset. The thus completed dataset then is used for the classifier’s construction. When learning a selective classifier, also the selection of appropriate features is based upon the completed data. The resulting classifier, however, is likely to be used in the original real-life setting where it is again confronted with missing values. By means of a real-life dataset in the field of oesophageal cancer that includes a relatively large number of missing values, we argue that especially the wrapper approach to feature selection may result in classifiers that are too selective for such a setting and that, in fact, some redundancy is required to arrive at a reasonable classification accuracy in practice.