In this article, we apply the “EM” algorithm to learning the parameters of a Markov chain mixture model for clus-tering navigation sessions on a Web site. Our main con-tribution is to deduce the model's initial parameters from clusters formed by a hierarchical clustering of a sample of sessions, whose dissimilarity matrix is computed by Dynamic Time Warping. The states of the Markov chains are the neurons of a Kohonen Self Organizing Map, which displays the site as it is seen by the users and also clusters its pages (one neuron corresponding to a cluster of pages). This technique for clustering sessions has been validated on a set of semi-artificial data and the results are excellent. Finally, we tested several criteria for the determination of the optimal number of clusters and con-cluded that the Akaike Information Criterion was best suited to this problem.
