Bayesian collaborative predictors for general user modeling tasks

Abstract

Collaborative approach is of crucial importance in user modeling to improve the individual prediction performance when only insufficient amount of data are available for each user. Existing methods such as collaborative filtering or multitask learning, however, have a limitation that they cannot readily incorporate a situation where individual tasks are required to model a complex dependency structure among the task-related variables, such as one by Bayesian networks. Motivated by this issue, we propose a general approach for collaboration which can be applied to Bayesian networks, based on a simple use of Bayesian principle. We demonstrate that the proposed method can improve both the prediction accuracy and its variance in many cases with insufficient data, in an experiment with a real-world dataset related to user modeling. © 2008 Springer-Verlag Berlin Heidelberg.