Concept drift is an important feature of real-world data streams that can make usual machine learning techniques rapidly become unsuitable. This paper addresses the problem of sudden concept drift in classification problems for which standard techniques may fail. To this end, support vector machines (SVMs) are automatically corrected to cope with a new suddenly drifted dataset. Results on real-world datasets with several types of sudden drift indicate that the method is able to correct the SVM in order to better classify the new data after the concept drift, using a correction based on the difference between the initial dataset and the new drifted dataset, even when the new dataset is small.
|Title of host publication||Proceedings de la 14 ème conférence Extraction et Gestion des Connaissances (EGC 2014)|
|Publication status||Published - 2014|