Statistical variable length Markov chains for the parameterization of stochastic user models from sparse data

This paper presents an algorithm for the parameterization of variable length Markov chains (VLMC). The disadvantages of the conventional algorithms especially in the application field of stochastic use models are discussed. The algorithm proposed in this paper eliminates these disadvantages by the usage of statistical methods to decide which states are accepted for the model. The benefit of this procedure is two-fold. First, the resulting models perform better in respect to prediction quality even they contain fewer states. Second, each state gets a more significant meaning. This makes the algorithm suitable for analyzing user-traces. An example for this application is described.