Bootstrapping and Decentralizing Recommender Systems

This thesis consists of three papers on recommender systems. The first paper addresses the problem of making decentralized recommendations using a peer-to-peer architecture. Collaborating recommender agents are connected into a network of neighbors that exchange user recommendations to find new items to recommend. We achieved a performance comparable to a centralized system. The second paper deals with the bootstrapping problem for centralized recommender systems. Most recommender systems learn from experience but initially there are no users and no rated items to learn from. To deal with this problem we have developed the Genesis method. The method bootstraps a recommender system with artificial user profiles sampled from a probabilistic model built from prior knowledge. The paper describes how this was done for a k-nearest neighbor recommender algorithm in the movie domain. We were able to improve the performance of a k-nearest neighbor algorithm for single predictions but not for rank ordered lists of recommendations. The third paper identifies a new domain for recommendations – product configuration – where new recommender algorithms are needed. A system that helps users configuring their own PCs is described. Recommendations and a cluster-based help system together with a rulebased configurator assist the users in selecting appropriate features or complete PC configurations. The configurator ensures that users cannot choose incompatible components while the recommender system adds social information based on what other users have chosen. This introduces new complexity in the recommendation process on how to combine the recommendations from the configurator and the recommender system. The paper proposes (1) three new recommender algorithms on how to make recommendations in the domain of product configuration, (2) a method for adding social recommendations to a rule-based configurator and (3) a method for applying the Genesis method in this domain. In this case the Genesis method is implemented by a Bayesian belief net that captures the designers’ prior knowledge on how to configure PCs. Then instances of complete configurations are sampled from the model and added to the recommender algorithm.