On stationarity in Internet measurements through an information-theoretic lens

Streaming change detection schemes are of great interest in the context of data warehouses, data cleaning systems, network traffic anomaly detection and other measurement scenarios. The goal of change detection systems is to determine whether fundamental characteristics of a data stream have changed, either temporally (with respect to prior data) or spatially (with respect to other streams). We present an application of an information-theoretic change detection scheme to determining stationarity in Internet measurements. Our experiments indicate that this generic scheme, despite being oblivious to the nature of the data, matches a more domain specific approach for this problem, and requires no domain knowledge to work effectively.