Projective inference in high-dimensional problems: prediction and feature selection (Q2188473)

The paper discusses predictive inference and feature selection for generalised linear models, when the data is highly dimensional but with a low number of samples, e.g. as in the case of microarray records. The article demonstrates that a two-step approach consisting of a reference model and a procedure for finding a minimal subset of features characteristic of the predictions is beneficial. Several techniques in this category are reviewed. A new clustered projection approach that is a combination of two other methods and has both speed and good accuracy is proposed. A novel LOO cross-validation method for evaluating the feature selection process is put forward. Also, a theorem for the conditions under which the projective approach is advantageous is proved. An R package is provided, where all the discussed methods are implemented. Several simulation studies performed and the application on five microarray data sets demonstrate also from the practical perspective the effectiveness of the approach.