Studying dependencies among Web traffic and link analysis data using perceptron

In this paper we try to experimentally check if it is possible to predict Web traffic data using only link analysis data. It is a natural continuation of Sydow, (2005), where correlations between link analysis and traffic data were measured. A perceptron is applied as an intelligent prediction module. The general conclusion is negative, i.e. dependencies mentioned above are too weak or too complex to be grasped by a perceptron. We also report some successful results concerning dependencies inside link analysis and inside traffic data. All the experiments are performed on intersection of 20-million sample from Polish Web graph and traffic data concerning 3 million Polish URLs.