TAKING ADVANTAGE OF UNLABELED DATA WITH THE ORDERED CLASSIFICATION ALGORITHM

We introduce a new method for improving poor perfor-mance of classi£ers due to a small training set. The Or-dered Classi£cation algorithm presented here incremen-tally increases the training set by adding unlabeled exam-ples. These unlabeled examples are selected by the algo-rithm accordingly to the con£dence level of the predictions made by an ensemble of classi£ers. The use of this con£-dence level measurement, which was inspired by the Query By Committee approach within the Active Learning set-ting, ensures that the algorithm incorporates the examples which are more likely to have the right classi£cation label assigned by the ensemble. Experimental results show that this algorithm effectively takes advantage of the unlabeled data yielding an error reduction of up to 78%. Giving that a very common scenario in classi£cation problems is the lack of a large enough training set, this algorithm provides a practical solution. KEY WORDS machine learning, unlabeled data, ensembles