Research On Recommendation Algorithm With Social Relations And Content Information In Social Networks

With the occurrence of Web2.0,social network services develop rapidly,and have become primary medium to spread or obtain information.As the explosive growth of users and information in social network services,the information overload problem appears.Thus,Social recommendation is proposed,which is a way to get personalized information.There are existing abundant heterogeneous information resources,such as social relation information and content information,which brings an opportunity to alleviate the data sparse and cold start problem from conventional collaborative filtering.At the same time,how to exploit those information effectively has become a big challenge.To address above issues,two novel social recommendation algorithms are proposed.One is a recommendation algorithm based on collaborative social deep leaming(CSDL),which is a hierarchical Bayesian hybrid model that effectively integrates action records data given by users,social information among items and content information of these items,and social information among users.Specifically,the social network of general users is introduced by adopting Jaccard's coefficient to compute a user similarity value,which can be utilized to constrain the taste difference between a general user and others.After that,deep learning technique is adopted to learn feature representation from sparse content information,and then factorizes the item's social relation matrix and the user's action record matrix simultaneously.Another one is a recommendation algorithm based on social collaborative deep learning(socialCDL),which is a effectively uniform model that confuses user's action records,social information among items and content information of these items.To be specific,item's social information is analyzed,which provides a physical interpretation of matrix factorization on item's social relation.After that,the item's social relation matrix and the user's action record matrix are factorized to achieve follower model and be-follower model.Then,to learn more effective feature from item content,a social factor regularized stacked denoising autoencoder(sfSDAE)model is utilized,which is an extension of conventional stacked denoising autoencoder(SDAE)model.The proposed two models take full advantage of these heterogeneous information,make them influence each other,learn from each other and remedy each other,which alleviates the data sparse and cold start problem from conventional collaborative filtering.Compared to three state-of-the-art algorithms on the Tencent Blog and Twitter datasets,the performances of two recommendation models have improved in terms of Recall and AP.