Although measurement-based real-time traffic classification has received considerable research attention, the timing constraints imposed by the high accuracy requirements and the learning phase of the algorithms employed still remain a challenge. In this paper we propose a measurement-based classification framework that exploits unsupervised learning to accurately categorise network anomalies to specific classes. We introduce the combinatorial use of two-class and multi-class unsupervised Support Vector Machines (SVM)s to first distinguish normal from anomalous traffic and to further classify the latter category to individual groups depending on the nature of the anomaly.