Sparse Approximations for High Fidelity Compression of Network Traffic Data

An important component of traffic analysis and network monitoring is the ability to correlate events across multiple data streams, from different sources and from different time periods. Storing such a large amount of data for visualizing traffic trends and for building prediction models of ``normal'' network traffic represents a great challenge because the data sets are enormous. In this paper we present the application and analysis of signal processing techniques for effective practical compression of network traffic data. We propose to use a sparse approximation of the network traffic data over a rich collection of natural building blocks, with several natural dictionaries drawn from the networking community's experience with traffic data. We observe that with such natural dictionaries, high fidelity compression of the original traffic data can be achieved such that even with a compression ratio of around 1:6, the compression error, in terms of the energy of the original signal lost, is less than 1%. We also observe that the sparse representations are stable over time, and that the stable components correspond to well-defined periodicities in network traffic.