Support vector machines (SVMs) are one of the most successful algorithms for classification. However, due to their space and time requirements, they are not suitable for on-line learning, that is, when presented with an endless stream of training observations.