Similarity-Based For Research On Friend Recommendation Algorithm

Friend recommendation as a core service of recommendation system plays a decisive role to increase user stickiness and expand user social circle.Most of the existing recommendation algorithms focus on the properties and connection intensity among nodes,and few of them consider the impact of social circles on friend recommendation from the global perspective.And the existing recommendation based on the social network topology puts emphasis on recommending the already known users and ignores the potential users who have the same interests.Recommendation based on the interests puts emphasis on recommending unfamiliar users,but the recommendation result is difficult to get the user's trust.At the same time,these two recommendation algorithms do not take user's behavior in the social network into account that greatly affects the accuracy,reliability and comprehensiveness of the recommended result.Secondly,the data format of social network is complex and changeable,and a recommendation algorithm does not solve the recommendation in all scenarios.To solve the above problems,the following research is done:1.This thesis designs a friend recommendation model based on social circle division.The model is suitable for the data sets with distinct characteristics of community and dense user labels.First,this model calculates the weight of the side between users according to the user label similarity,and the weight is fused to the Fast Unfolding algorithm.Then,the similarity of any two users in the same social circle is calculated based on the friend distribution.Then the similarity is sorted to generate top-N recommendation list.2.This thesis designs a friend recommendation integrating confidence in social networks.First,a fusion similarity calculation method is designed.The user's social similarity is calculated according to the common concern friends of the nodes in the social network.The keywords are extracted based on TF-IDF algorithm and the weight vectors of these keywords are calculated,then the keywords are used to express user's interest and calculate interest similarity among users.The best fusion parameter is found through experiments and the fusion similarity between two dimensions of social interaction and interest is calculated.Then,a confidence calculation method is designed in this thesis.The method comprehensively considers the user social network topology and social network behavior.This model measures the confidence factor of the relationship by using the number of common neighbors and calculates the confidence factor of behavior based on the user's interaction in the social network.The confidence is calculated by integrating the confidence factors of the behavior and relationship.The initial recommendation result is revised by use the confidence degree,so that the recommendation result has higher credibility and accuracy.The experiment is done through Sina micro-blog real data set.The experimental results show that the friend recommendation model based on the social circle can optimize the results of community classification,and can produce good recommendation results when the user tags are dense and other data is sparse.A friend recommendation integrating confidence can produce good recommendation results on tag sparse data sets.