Research And Implementation Of Recommendation Algorithm Based On Network Embedding

With the rapid development of the Internet and information technology,information overload problem has become increasingly serious.Personalized recommendation is an effective way to solve information overload,it has been widely used in news,e-commerce and other fields.As the core of personalized recommendation,recommendation algorithm has become a research hotspot in recent years.Collaborative filtering algorithm is currently the most widely used recommendation algorithm,but there are still some problems in this algorithm.It only considers two types of entities,user and item,ignoring contextual information such as user attributes,item attributes,time,user behavior patterns,etc.So the recommendation results are often unsatisfactory.In addition,while data scale continues to expand,suchalgorithms will face serious scalability issues.The network embedding technology can effectively solve the scalability problem of the recommendation algorithm,and it provides a new idea for the research of the recommendation algorithm.Based on the network embedding technology,this thesis deeply discusses and studies the recommendation algorithm considering contextual information.According to different contextual information fusion methods,we propose two recommendation algorithms based on network representation learning:User-Item Graph based Network Embedding Recommendation(UIGNER)and Meta-Path based Network Embedding Recommendation(MPNER).Both algorithms consider four kinds of contextual information:user attributes,item attributes,time,and user behavior patterns.UIGNER uses the contextual information to construct the user-item graph and reasonably quantifies the weights of the three types of edges between the users and items.It applies network embedding methods to obtain the vector representation of users and items,and then mines the correlation between users and items and recommends items through the correlation.MPNER converts contextual information into entity nodes and uses these entity nodes to coinstruct corresponding heterogeneous information networks,and applies network embedding method to learn user feature vectors on different meta-paths.Then,the Bayesian Personalized Ranking algorithm is used to calculate the user similarity weight matrix on different meta-paths.The similarity weight matrix is used to fuse the user similarity on different meta-paths,and finally the recommendation is based on the fusion user similarity results.According to experimental verification of the MovieLens public dataset and the mobile company OTT TV real dataset,these two algorithms have better recommendations than other classic recommendation algorithms.First,this thesis introduces the background and significance of the topic,and analyzes some state-of-the art recommendation algorithms and network embedding algorithms.Then,it analyzes the contextual information of the two datasets in movie recommendation scenario and Internet TV program recommendation scenario and does some data processing works,such as data crawling,data cleaning and data formatting.On this basis,the thesis considers four kinds of contextual information:user attributes,item attributes,time and user behavior patterns when designing algorithm models.It proposes two kinds of recommendation algorithms based on network embedding.Then the experiments on recommended quality and parameter adjustment are designed for the two algorithms.The experimental results prove the feasibility and effectiveness of the proposed algorithms.Finally,the thesis summarizes the author's work and formulates the feature research direction.