Decomposition of network traffic of dialup access into trend and periodic components based on nonstationary Bayesian model

A new model to decompose a network traffic time series into its components, such as trend, the effect of days in a week, and variations in a day, is proposed. The logarithm of dialup access traffic at Hosei University, which was averaged over an hour, is used as the network traffic time series. The idea of the model is originated from a nonstationary Bayesian model that decomposes a monthly tune series into trend, seasonally, and trading day effects in a month. In our approach, the components which appear in the model are defined as follows: there are two types of trend component which are smooth over several days and several hours; the effect of days in a week is a periodicity in a week; and the variation in a day is a periodicity in a day. We conclude that a model of one combination of the components is suitable for analysis, through an evaluation that uses ABIC for several combinations of the components.