Recommendation systems are important in social networks that allow the injection of user-generated content and let users indicate their preferences towards the content introduced by others. Considering the increase of usage of these collaborative systems, it seems only a matter of time before the current centralized systems will be replaced by decentralized solutions. However, current collaborative filtering systems assume that recommendations can be based on the entire data collection in the network. This work evaluates the performance of user-based collaborative filtering systems when only partial knowledge about the network is available at an end-user’s computer. We propose a utility model that combines three important aspects of network users (similarity, confidence and usefulness) in order to create a semantic overlay network optimized for autonomous content recommendations. We compare different similarity functions on the most common dataset in collaborative filtering and we show the influence of the confidence and usefulness parameters on both dense and sparse data. We find that the commonly used similarity function results in sub-optimal performance when used as updating criterion for locally stored rating profiles. We show that taking into account the level of confidence in the computed similarity can greatly improve recommendation accuracy, especially when a small user neighborhood is selected. Also, conventional methods select many users that cannot contribute to the recommendation, because they have rated too few items. The usefulness parameter that we introduce compensates for this problem, so that even a small local cache in very sparse data provides valuable recommendations.

Workshop on large scale distributed systems for information retrieval
Database Architectures

Clements, M, de Vries, A.P, Pouwelse, J.A, Wang, J, & Reinders, M.J.T. (2007). Evaluation of Neighbourhood Selection Methods in Decentralized Recommendation Systems. In Workshop on Large Scale Distributed Systems for Information Retrieval (LSDS-IR) (pp. 38–45). A.C.M.