Trust networks for recommender systems

Recommender systems use information about their user’s profiles and relationships to suggest items that might be of interest to them. Recommenders that incorporate a social trust network among their users have the potential to make more personalized recommendations compared to traditional systems, provided they succeed in utilizing the additional (dis)trust information to their advantage. Such trust-enhanced recommenders consist of two main components: recommendation technologies and trust metrics (techniques which aim to estimate the trust between two unknown users.) We introduce a new bilattice-based model that considers trust and distrust as two different but dependent components, and study the accompanying trust metrics. Two of their key building blocks are trust propagation and aggregation. If user a wants to form an opinion about an unknown user x, a can contact one of his acquaintances, who can contact another one, etc., until a user is reached who is connected with x (propagation). Since a will often contact several persons, one also needs a mechanism to combine the trust scores that result from several propagation paths (aggregation). We introduce new fuzzy logic propagation operators and focus on the potential of OWA strategies and the effect of knowledge defects. Our experiments demonstrate that propagators that actively incorporate distrust are more accurate than standard approaches, and that new aggregators result in better predictions than purely bilattice-based operators. In the second part of the dissertation, we focus on the application of trust networks in recommender systems. After the introduction of a new detection measure for controversial items, we show that trust-based approaches are more effective than baselines. We also propose a new algorithm that achieves an immediate high coverage while the accuracy remains adequate. Furthermore, we also provide the first experimental study on the potential of distrust in a memory-based collaborative filtering recommendation process. Finally, we also study the user cold start problem; we propose to identify key figures in the network, and to suggest them as possible connection points for newcomers. Our experiments show that it is much more beneficial for a new user to connect to an identified key figure instead of making random connections.

Grantor dc:publisher

Ghent University. Faculty of Sciences

Year dc:date

2010

Author dc:creator

• Victor, Patricia

Contributors dc:contributor

• De Cock, Martine
• Cornelis, Chris

• Technology and Engineering

Statement dc:rights

• info:eu-repo/semantics/openAccess

Language dc:language

eng

Identifier

https://biblio.ugent.be/publication/986279
https://biblio.ugent.be/publication/986279/file/4335314

OAI identifier oai:identifier

oai:archive.ugent.be:986279