Book Recommendation Based On Heterogeneous Network Embedding

With the continuous enrichment of book resources,it is difficult for users to obtain books which they are interested in,therefore book recommendation techniques are followed.Predict books that users are interested in by using user information,book information,and user historical behavior information,so that users can obtain related books quickly and accurately and save a lot of time.In addition,high-quality book recommendation methods can also help the book sales platform enhance its ability to accurate recommendation,thereby enhancing its core competitive advantage.The book recommendation mainly includes the following three methods:recommendation based on content,collaborative filtering and association rules.These three book recommendation methods are easy to understand,the requirements for input data are simple,and the book recommendation has achieved certain effects.However,the information used is not sufficient,these three book recommendation methods are only based on the content,the relationship between book and user,the relationship between books respectively.The relationships between books,authors,categories and other more characteristic information are often ignored.The recommended books are very similar in content,lack of diversity,and various types of recommendations.Recommending books of various categories can provide more choices for users and satisfy the users with wide interests.Therefore,this paper intends to integrate more book feature information in the recommendation process,thereby improving the accuracy of the recommendation,and recommending a variety of books related more than similar.Network embedding method can effectively solve the problems above and integrate the nodes and the relationship in the network,so that we can the full use of books,users,categories,publishers,and the relationships between them.The nodes in the network can be represented as dense low-dimensional vectors,thereby efficiently calculating the connections between nodes in the network and improving the accuracy of book recommendation.Therefore,this paper conducts experiments on the data of the public Amazon ecommerce transaction records based on network embedding method.Firstly,extract the book features and define the multi-dimensional relationship between the features in order to construct a book heterogeneous network.Then integrate the book features based on heterogeneous network embedding so that a book can be represented as a semantic vector.From the perspective of semantic relevance,recommend books by calculating the cosine similarity between books.In terms of recommendation accuracy,the two indicators MAE and RMSE of book recommendation based on heterogeneous network embedding are lower than the book recommendation method based on collaborative filtering,RMSE and MAE respectively reduced by 19.52%,20.51% mostly,which shows book recommendation based on heterogeneous network embedding is more accurate.In terms of recommendation relevance,we analyzed from two aspects of category diversity and content diversity.The result shows that the book recommendation methods based on content and collaborative filtering emphasize similarity,and the book recommendation based on heterogeneous network embedding emphasizes relevance.The methods based on heterogeneous network embedding recommend books with more categories and content.In summary,we verify the feasibility and effectiveness of the heterogeneous network embedding in the book recommendation,and we explored the contribution of different features on recommendation.