Recommender systems are widely used to achieve a constantly growing variety of services. Alongside with social networks, recommender systems that take into account friendship or trust between users have emerged. In this thesis, we propose an evolution of trust-based recommender systems adapted to decentralized architectures that can be deployed on top of existing social networks. Users profiles are stored locally and are exchanged with a limited, user-defined, list of trusted users. Our approach takes into account friends' similarity and propagates recommendation to direct friends in the social network in order to prevent ratings from being globally known. Moreover, the computational complexity is reduced since calculations are performed on a limited dataset, restricted to the user's neighborhood. On top of this propagation, our approach investigates several aspects. Our system computes and returns to the final user a confidence on the recommendation. It allows the user to tune his/her choice from the recommended products. Confidence takes into account friends' recommendations variance, their number, similarity and freshness of the recommendations. We also propose several heuristics that take into account peer-to-peer constraints, especially regarding network flooding. We show that those heuristics decrease network resources consumption without sacrificing accuracy and coverage. We propose default scoring strategies that are compatible with our constraints. We have implemented and compared our approach with existing ones, using multiple datasets, such as Epinions and Flixster. We show that local information with default scoring strategies are sufficient to cover more users than classical collaborative filtering and trust-based recommender systems. Regarding accuracy, our approach performs better than others, especially for cold start users, even if using less information.
