Research And Application Of Friend Recommendation Algorithm Based On Social Network

The research and application of friend recommendation algorithms is a hot topic relate to recommendation based on social network. With the social network service was widely used in many fields, recommend friends to users by digging the latent interesting of users has been attached more and more attentions. Traditional friend recommendation algorithms used the number of mutual friends between users to recommend friend for them, the more mutual friends, the possibility was recommended is higher. However, this kind of algorithms have’Cold Start’problem, so traditional algorithms are unsuitable for recommend friends to new users, because the number of new users’friends is low. Besides the algorithm based on mutual friends, item-based collaborative filtering algorithm also be applied to friend recommendation, but the problem of sparse rating data affecting the accuracy of recommendation. In order to solve the problems of traditional recommendation algorithms, we studied the friend recommendation algorithms, and proposed improved algorithms which have been applied to a practical application and have a good effect.First, two recommendation algorithms based on common interests was proposed in this paper. The first algorithm recommend friends uses nearest-neighbor model based on item to recommend friend for users, it build the user-item rating matrix depend on following relationship between users, and optimized nearest-neighbor model using normalized method. This method could solve the’Cold Start’problem, and improve the coverage scale and diversity of recommendation. The second algorithm is the improved algorithm based on LDA, the main idea of this algorithm is that extract the keywords or tags about user’s interest form the information which be broadcasted, read or forwarded in social network. The LDA topic model extract the interest topic of users, and calculate the user’s similarity on topic level. It defined new regularization terms based on user’s neighborhood and integrated these terms into matrix factorization model, a matrix factorization recommendation model based on LDA was built at last. The experimental result shows that this model could improve the accuracy rate of recommendation.Second, because of the matrix factorization model has excessive dependence on use-item rating matrix and not take full advantage of the structured information of the social network. We proposed a matrix factorization recommendation model based on social network regularization. In this model, we use different similarity function to calculate users’similarity for each kind of structural feature, and defined regularization term based on different users’similarity. These regularization terms could minimize the distance between users’latent feature vector in matrix factorization model, and also could alleviate the over-fitting problem for training model. The improved model integrated the users’relationship in social network into matrix factorization model as a kind of auxiliary information. This method was validated using Tencent weibo dataset, and the result shows that this method has a higher Mean Average Precision than traditional recommendation methods.At last, we design and implement a simple recommendation system using recommendation algorithms we proposed in this paper, and applied this recommendation system onto www.woao.com. This recommendation system could recommend friends for new users, recommend friends for users based on mutual friends and common interests.