Randomly perturbed random forests

In supervised classification, a change in the distribution of a single feature, a combination of features, or the class boundaries, may be observed between the training and the test set. This situation is known as dataset shift. As a result, in real data applications, the common assumption that the training and testing data follow the same distribution is often violated. In order to address dataset shift we propose to randomly introduce more variability in the training set by sketching the input data matrix resorting to random projections of units. We then modify the random forests algorithm to involve sketched, rather than bootstrapped, versions of the original data. Results on real data show that perturbing the training data via matrix sketching improves the prediction accuracy of test units that have a different distribution in terms of variance structure.