The objective of recommender systems is to estimate the unknown ratings. This paper presents an efficient method to generate the self-similarity rating matrix for recommender systems. We show that the rating behavior of users is statistically self-similar that none of the commonly used recommender system models is able to detect this fractal behavior. This behavior can be used to predict the unknown ratings. The experimental results showed that the proposed method obtains similar accuracy in comparison to the traditional recommender system method with much less computational cost.