Masashi Nakatomi

Bayesian Collaborative Predictors for General User Modeling Tasks

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.

Context gallery: a service-oriented framework to facilitate context information sharing

Context-aware computing enables the users to seamlessly manipulate information through the relevant contexts, however, it is problematic for application developers to search for the particular contexts they could use, and to determine how reliable the contexts are and how the contexts are depending on each other. We propose a XML-based context information syndication format called “context summary”, and its browsing web-based service called “context gallery”. We demonstrate several novel interactive systems leveraging various context information, and we then discuss how these applications are loosely interconnected with each other through contexts on our framework.