Research On Friends Recommendation For Social Networks

As the product of development of Web2.0technology and social media, socialnetworks changes the mode of information fusion and information transmission. Friendsrecommendation is a key technology part of social networks platform. It can not onlyrecommended users with new friends, improve user’s loyalty to the social platform, butalso create profit for manufacturers by mining user’s preference to recommend serviceor products.This article summarize state-of-art personalized recommendation technology onsocial networks, especially friends recommendation research. Based on link predictionalgorithms of complex networks, this paper purpose a new implementation of localrandom walk model and prove it with comparison test. This model can reduce thecomputation cost of full graph random walk while keep the prediction accuracy, thenmeet the needs of online real-time friends recommendation. We chose to ignore nodeswhich far away from the start based on local principle, and continue to add resource toeach new walk for better personalized recommendation.There are not only large scale social relationship data on online social networks,but also rich heterogeneous information contains user features and user activities etc. Sohow to deal with these large-scale, sparse,imbalance and heterogeneous data is a puzzlefor certain friends recommendation models. It is known that matrix factorization modelhas achieve good results for several data mining tasks for it’s ability to handlelarge-scale and sparse data, we also purpose a improved matrix factorization model tointegrate the user social relationship informatio n and user profile, tag etc. This modelcan comprehensive utilize the social networks’ rich data, and improve therecommendation accuracy.Experiments based on Tecent microblog open data set show our models efficienceand effectivity. This dataset contains user social relationships, activities, profiles, tagsetc. The results show that SLRWR model is effective with at least6.4%in accuracyimprovement and4.6%in ranking accuracy improvement compared with test models,and the WSVD series models are much more physical acceptable with full usage ofdiverse information, especially when utilizing the social relationship regularization with16.9%accuracy improvement compared with native models.