Towards Decentralized Recommender Systems
Cai-Nicolas Ziegler · FreiDok plus (Universitätsbibliothek Freiburg) · 2005
Automated recommender systems make product suggestions that are tailored to the human user's individual needs and represent powerful means to combat information glut. However, their practical applicability has been largely confined to scenarios where all information relevant for recommendation making is kept in one single, authoritative node. Recently, novel distributed infrastructures are emerging, e.g., peer-to-peer and ad-hoc networks, the Semantic Web, the Grid, etc., and supersede classical client/server approaches in many respects. These infrastructures could likewise benefit from recommender system services, leading to a paradigm shift towards decentralized recommender systems. In this thesis, we investigate the challenges that decentralized recommender systems bring up and propose diverse techniques in order to cope with those particular issues. The spectrum of methods proposed ranges from the employment of product classification taxonomies as powerful background knowledge, alleviating the sparsity problem, to trust propagation mechanisms designed to address the scalability issue. Empirical investigations on the correlation of interpersonal trust and interest similarity provide the component glue that melds these results together and renders the eventual creation of a decentralized recommender framework feasible. While these building bricks, namely taxonomy-driven filtering, topic diversification, and the Appleseed trust metric, are vital for the conception of our trust-based decentralized recommender, they are also valuable contributions in their own right, addressing issues not only confined to the universe of decentralized recommender systems.