Landmarking for Meta-Learning using RapidMiner
In machine learning, picking the optimal classifier for a given problem is a challenging task. A recent research field called meta-learning automates this procedure by using a meta-classifier in order to predict the best classifier for a given dataset. Using regression techniques, even a ranking of preferred learning algorithms can be determined. However, all methods are based on a prior extraction of meta-features from datasets. Landmarking is a recent method of computing meta-features, which uses the accuracies of some simple classifiers as characteristics of a dataset. Considered as the first meta-learning step in RapidMiner, a new operator called landmarking has been developed. Evaluations based on 90 datasets, mainly from the UCI repository, show that the landmarking features from the proposed operator are useful for predicting classifiers' accuracies based on regression.