Data mining network traffic

In this paper we present a novel approach to network traffic analysis. In particular, we show how to determine which statistical traffic descriptors are most pertinent in predicting important network performance metrics such as packet loss rate, based on empirical data. In addition, we reveal the relationship between the pertinent traffic descriptors and packet loss rate via fuzzy if-then rules. The principal finding of this paper is that descriptors that quantify intermittency such as the generalized fractal dimensions D-1, D-2 and D-3 or parameter that quantify variability such as the Holder exponent hi are better indicators of packet loss rate than the more commonly used Hurst parameter H and tail exponent a of a long-range dependent and heavy-tail random variable, respectively. A simple fuzzy inference system that incorporates rules generated from these traffic descriptors was able to predict the packet loss rate reasonably well, verifying the above claim.