To improve Bayesian Network Learner Modelling thanks to Multinet

Bayesian Network (BN) are often used for student modelling but some problems remain, particularly the question of the structure of the BN. The key idea of this paper is that the structure of a BN-based Student Model (BNb-SM) depends on the level of expertise of the student. Therefore, a model should be constituted with several concurrent BNs (same nodes, different structures), instead of a single one, as it is usually the case. We present a conceptual model, the multinet, that allows to take into account different BNs. A multinet is a probabilistic graphical knowledge representation that can be seen as a BN mixture. We present both theoretical and experimental results obtained with real student's data. These results give strong evidence in favour of our approach by showing that there is a correlation between the student's levels of expertise and the Bayesian networks which fit their interactions best.