Abstract: Collaborative filtering is one of prevalent successful methods of recommender system. In this study, a prediction model for homophily clustering of users has been built to improve the collaborative filtering recommender system. The general framework mainly consists of two phases: Firstly, detect communities in homophily networks by using Partitioning Around Medoids (PAM) clustering algorithm. Secondly, building naive bayes model by calculating the conditional probability foruser’s demography attributes. The experiments have been applied on two real world datasets 100 K and 1 M that published by grouplens. Finally, precision, recall and F-measure metrics have been used to evaluate the top-N recommendation lists. The empirical results can provide a recommendation in a best manner also, the results have compared with other research study.