Controlling Consistency in Top-N Recommender Systems

Recommender systems have become essential navigational tools for users to surf through vast on-line catalogs. However, recommender algorithms are often tuned to improve accuracy, without paying any attention to the consistency of the recommendations when small changes happen to the user profile or to the model. Consistency of recommendations is closely related with user satisfaction and trust. In this work we analyze how small changes in either the user profile or the recommender model may affect the consistency of Top-N recommendation systems. We also design two mechanisms able to promote consistency without degrading accuracy and novelty of recommendations. Finally, we investigate the consistency of Top-N recommendation algorithms over time by analyzing real data from a production IPTV recommender system.