ipred : Improved Predictors

In classification problems, there are several attempts to create rules which assign future observations to certain classes. Common methods are for example linear discriminant analysis or classification trees. Recent developments lead to substantial reduction of misclassification error in many applications. Bootstrap aggregation (“bagging”, Breiman, 1996a) combines classifiers trained on bootstrap samples of the original data. Another approach is indirect classification, which incorporates a priori knowledge into a classification rule (Hand et al., 2001). Since the misclassification error is a criterion to assess the classification techniques, its estimation is of main importance. A nearly unbiased but highly variable estimator can be calculated by cross validation. Efron and Tibshirani (1997) discuss bootstrap estimates of misclassification error. As a by-product of bagging, Breiman (1996b) proposes the out-of-bag estimator. However, the calculation of the desired classification models and their misclassification errors is often aggravated by different and specialized interfaces of the various procedures. We propose the ipred package as a first attempt to create a unified interface for improved predictors and various error rate estimators. In the following we demonstrate the functionality of the package in the example of glaucoma classification. We start with an overview