SVM learning of IP address structure for latency prediction

We examine the ability to exploit the hierarchical structure of Internet addresses in order to endow network agents with predictive capabilities. Specifically, we consider Support Vector Machines (SVMs) for prediction of round-trip latency to random network destinations the agent has not previously interacted with. We use kernel functions to transform the structured, yet fragmented and discontinuous, IP address space into a feature space amenable to SVMs. Our SVM approach is accurate, fast, suitable to on-line learning and generalizes well. SVM regression on a large, randomly collected data set of 30,000 Internet latencies yields a mean prediction error of 25ms using only 20% of the samples for training. Our results are promising for equipping end-nodes with intelligence for service selection, user-directed routing, resource scheduling and network inference. Finally, feature selection analysis finds that the eight most significant IP address bits provide surprisingly strong discriminative power.