Performance and flexibility of stereotype-based user models

Since it was first proposed by Rich in 1979, stereotype-based user modelling has been applied numerous times in recommender systems. The primary motivation for stereotyping in user modelling is the new user problem — a purely individualised user model cannot be constructed for a user until he has provided some ratings of items. By appealing to a pool of manually-constructed stereotypes, each one representing the interests of a set of users with common socio-demographic attributes, and eliciting enough information from the user to match him to a set of stereotypes, the matching stereotypes can be combined and used to recommend items to the user. Many claims have been made in the literature regarding the efficacy of stereotyping but these have not been substantiated empirically. This thesis describes the first comprehensive empirical investigation into the recommendation performance of stereotype-based user models and details an approach to training stereotype-based user models automatically. The recommendations provided by stereotype-based user models are directly compared to those provided by individualised models and it is shown that the performance levels of the two approaches are comparable on average. However, stereotype-based user modelling does not work well for all users (whether new or known). This thesis demonstrates that individual stereotype performance is meaningless in isolation but that it is the interaction between the recommendation lists of stereotypes that determines overall user model performance. Concepts from hybrid modelling are used to explain the variation in performance of stereotype-based user models across the user population. In addition, robustness to abrupt changes in a user’s information requirements, not a major concern in current user modelling literature, is highlighted as an important issue for users in agile team settings. Experimental results show that stereotypes can be reassigned, added and removed without a significant reduction in recommendation performance.