Research On Embedding Learning Technology And Recommendation Algorithm Based On Social Network

The advent of the Internet era has given birth to a wide variety of social applications that have greatly improved people's living.And the scale of these social software is expanding with the development of information technology,forming many large-scale social networks such as QQ,maimai,Weibo,etc.,and accumulating a large amount of meaningful data.How to mine the potential social and commercial value through some social network analyses in these massive data is an urgent problem.Social network embedding,namely,embedding social network nodes into a low-dimensional space,is the foundation of social network analysis.Although many ex-isting methods attempt to address this task,most of them only consider the shallow relationship between two nodes,which ignore capturing multiple and semantic-rich so-cial relationships between users.To this end,this paper defines such relationships as multi-path relationships,and proposes a multi-path relationship preserved social net-work embedding method,which is based on the recurrent neural network framework that incorporates both network structure and node profile information.Meanwhile,social networks also generate a new application scenario on the rec-ommendation system,namely,the social recommendation system.Combined with the massive data generated by social networks and commodities,it's greatly helpful for commodity promotion and users to find the interesting items.However,most existing social recommendation methods only rely on the direct friends of the user or the pre-set meta-path mode,and fail to make full use of the category preference information and rating preference information on the social recommendation system.Based on this,this paper utilizes heterogeneous information network to model the social recommendation system,and proposes a heterogeneous information network embedding based social recommendation algorithm.Specifically,the contents and achievements of this paper are as follows:1)A multi-path relationship preserved social network embedding method is pro-posed.This paper first utilizes random walks to explore the multiple social relationship paths between nodes.Then,a new recurrent unit called bi-directional multi-path re-lationship unit is proposed to better capture the properties of multi-path relationships.Finally,two objective functions are designed to seamlessly integrate social network structure and node profile information into node representation.The experimental re-sults on two real-world networks show that the proposed algorithm outperforms the state-of-the-art baselines on node classification task and link prediction task.2)A heterogeneous information network embedding based social recommenda-tion method is proposed.In order to make full use of the node representation obtained by network embedding learning,this paper introduces the pre-training and fine-tuning approach into the recommendation algorithm.In the pre-training stage,transformer encoder is used to extract the features in the sequences obtained by the weighted ran-dom walks,and two objective functions are designed to make the model capture the category preference information and rating preference information respectively.In the fine-tuning stage,this paper adds a simple linear combination to the pre-training model and obtains an end-to-end rating prediction model by fine-turning the known rating data.Finally,the experimental results on two real social recommendation datasets show that the proposed model outperforms the baselines and can better alleviate the cold start problem.