Recommender systems use data on past user preferences to predict possible future likes and interests. This paper shows an algorithm that uses diversity instead of similarity to improve accuracy without using semantic or context-specific information.