Recently, recommender systems have been a vital part of social network applications to help users make the best choices and expand their heuristic experiences. Friend recommender is a popular service that has been widely developed and focuses mainly on relationships and user interests. Traditional friend recommender systems suffer from some shortcomings that hamper their effectiveness. Hybrid techniques to recommender systems can significantly improve the overall quality of related applications. In this research, a hybrid cumulative knowledge framework (HCKF) has been designed for friend recommendation in social networks. HCKF performs a huge amount of precise evaluations on data, including pre-processing on the data, data preparation, optimal features selection and user clustering in the background; therefore, friend suggestions can be provided at an acceptable performance, precision and speed. HCKF solves the cold start challenge of new users, increases the accuracy of providing attractive friends to users and also minimizes the error rate. The experimental results demonstrate the advantages and effectiveness of the proposed framework in social networks. After simulating HCKF and calculating the error rate and accuracy, the final precision rate of the friend suggestion to the newly logged-in user is equal to 90\%. Moreover, the average accuracy in the proposed framework is equal to 0.0197, which has improved significantly compared to other related methods.

Nowadays, social networks are becoming more popular, so the number of their users and their information is growing accordingly. Therefore, we need a recommender system that uses all kinds of available information to create highly accurate recommendations. Regarding the general structure of these recommender systems, one criterion is first chosen to calculate the similarity between users and then people who are assumed to have great similarity are proposed to each other as friend. These similar criteria can calculate users’ similarity with regard to topologica l structure and some properties of graph vertices. In this paper, the properties that are required for clustering are extracted from users’ profile. Finally, by combining the similarity criteria of mean measure of divergence (MMD), cosine, and Katz, different aspects of the problem including graph topology, frequency of user interaction with each other, and normalization of the same scoring method are considered.