Research On Deep Learning Recommendation Algorithm Based On Heterogeneous Information Network

With the advent of the Internet era,users can obtain a lot of information on the Internet,which provides users with great convenience.At the same time,users are also worried about how to quickly and effectively obtain useful information from the vast ocean of knowledge,which can not satisfy users of individual needs.The emergence of recommendation systems can just meet the needs of users,and it can provide users with personalized recommendations.The recommendation systems can change all aspects of people's lives,and they play an indispensable role in people's lives.At present,collaborative filtering algorithm is the most widely used recommendation algorithm.It mines user's preferences through user's historical interaction records and it recommends similar items to the user.Although traditional recommendation algorithm has achieved great recommendation effect,however,the information considered is not comprehensive,and the fusion of other auxiliary information is not considered.Heterogeneous information network can just use various types of nodes and the relationships between nodes to fuse richer information and make recommendations based on meta-paths.In addition,deep learning can automatically learn the latent features of useful data in a nonlinear way.Therefore,this thesis adopts deep learning techniques to study recommendation algorithms in heterogeneous information networks from the perspective of meta-paths.The main contributions of this thesis are as follows:(1)Existing heterogeneous information network based recommendations used to extract meta-path semantic information through max pooling.They do not consider the influence of integral features in meta-path semantics and the problem of feature redundancy.To solve these problems,this thesis proposes a recommendation model named MICA,which fuses meta-path and improved collaborative attention mechanism.First,MICA uses a collaborative attention mechanism based on k-max pooling to learn the collaborative attention embedding representations of users and items in a deeper level,respectively.It alleviates the problem of feature losses caused by max pooling in the collaborative attention mechanism.Second,based on the significant feature and integral feature,the meta-path context embedding representation is obtained through the attention mechanism,a mechanism to learn important feature of metapath,to retain effective meta-path information.Last,to reduce redundant information,the user and item collaborative attention embedding representations and meta-path context embedding representations are fused to input into attention mechanism for top-N recommendations.(2)Users and items learn different latent features under different symmetric meta-paths.In view of the differences in semantic information contained in them,this thesis proposes a recommendation model named CBMA,which combines balanced perception matrix and attention.For latent features learned by users and items under different meta-paths,the model uses the learned balance coefficient to balance the features obtained under each meta-path,which alleviates the problem of feature information losses.When fusing user and item features,since meta-path has the problem of feature overlap when learning features,an attention mechanism is introduced to reduce the redundant information of features.In this thesis,this thesis studies deep learning recommendation algorithm based on heterogeneous information network.In order to learn feature information based on meta-path more fully,this thesis propose two recommendation models,a recommendation model is that combines meta-path and improved collaborative attention,and another recommendation model is that combines balanced perception matrix and attention.The proposed models are better than other models,and they can better improve the recommendation performance,which verifies the effectiveness of the proposed model.