Testing rank and compare SVM on simulated patterns
compare.Rd50, ..., 800 labeled pairs were generated according to the squared L1, L2, and Linf norms, with input features in [-3, 3]. Labels are -1 when there is no significant difference between the ranks of Xt and Xtp, and 1 when the rank of Xtp is significantly higher than the rank of Xt. SVMrank ignores the pairs labeled -1, but SVMcompare uses them to learn a more accurate predictive model, as shown by the test error and level curves of the learned functions.
Usage
data(compare)Format
List of 4 data.frames: error contains the test error of the learned models, bayes contains the Bayes classification error of the latent ranking function applied to the test data, rank contains the ranking functions evaluated on a grid of input points, train.pairs contains the data that were used to train the algorithms.