This paper investigates a number of techniques for calibration of the output of a Support Vector Machine in order to provide a posterior probability P(target class | instance). Five basic calibration techniques are combined with five ways of correcting the SVM scores on the training set. The calibration techniques used are addition of a simple ramp function, allocation of a Gaussian density, fitting of a sigmoid to the output and two binning techniques. The correction techniques include three methods that are based on recent theoretical advances in leave-one-out estimators and two that are variants of hold-out validation set. This leads us to thirty different settings (including calibration on uncorrected scores). All thirty methods are evaluated for two linear SVMs (one with linear and one with quadratic penalty) and for the ridge regression model (regularisation network) on three categories of the Reuters Newswires benchmark and the WebKB dataset. The performance of these methods are compared to both the probabilities generated by a naive Bayes classifier as well as a calibrated centroid classifier. The main conclusions of this research are: (i) simple calibrators such as ramp and sigmoids perform remarkably well, (ii) score correctors using leave-one-out techniques can perform better than those using validation sets, however, cross-validation methods allow more reliable estimation of test error from the training data.
Search
Reviews
[Write a review of this article]
Find related articles from these CiteULike users
