Semantic clustering of information systems' users with stochastic techniques

We use a Markovian model to capture the habitual user profiles of an information access system. In this model, the general as well as the individual for each user, profile is captured in the form of a Markovian process where the states are the keywords asked to the system by the users and a transition from state to state corresponds to the order theses keywords appeared in the queries. Under this model the probabilistic locality of the Markovian state space translates to semantical locality of the corresponding keywords in a way that a clustering of the Markovian state space corresponds to a semantic clustering of the keyword space. Since the states represent keywords asked by the users, the state space can grow very large, but at the same time it is partitioned into disjoint subsets such that strong interactions among the states of the same subset exists but weak interactions among states of different subsets. We exploit this structure to effectively cluster the large state space and reveal the corresponding semantic keyword clusters. We then define a semantic distance between the various user profiles that can be used to cluster the user space on the basis of keyword usage and keyword semantic relevance. The resulting clustering achieves high independence from the row data. Users for e.g. that never asked a common keyword may end up very close to each other if their keywords were asked together by many other users.