Research On Heterogeneous Network Based News Recommendation Method

News is one of the important research objects in the field of natural language processing,and mining content-related news from a large amount of news for recommendation can help improve users’ reading experience or be used for related news content analysis.When modeling features of news,traditional text encoders have difficulty in preserving the rich semantic information and syntactic structure in long documents,so graph representation of documents is proposed to discover the potential semantic structure of long-form documents.However,news documents often contain complex semantic interaction information,and it is challenging to model the diverse semantic information into a network and learn effective representation from it for related news recommendation systems.Heterogeneous information networks can model complex contextual information,and their performance is not only excellent in feature modeling,but also widely used in recommender systems.Therefore,the Heterogeneous Network Based News Recommendation Method(HNNR)is proposed for related news recommendation tasks.HNNR constructs the news corpus as a heterogeneous information network with multiple types of nodes and edges,and effectively uses the multiple interactions of news texts to learn news features for Top-K related news recommendation.Specifically,HNNR unsupervisedly extracts co-occurrence and association information from the news corpus as interaction relations to model news interaction networks.To construct meta-path-based contexts,a combination of a weighted random walk algorithm and a meta-path-based random walk algorithm is used to preferentially sample path instances that characterize important interaction information.Through a self-attentive mechanism and a multi-view strategy,HNNR is able to learn effective representations of news under multiple interaction relations,enabling an unsupervised semantic matching-based recommendation method.To verify the effectiveness of the HNNR method,the performance is tested on two real news datasets,CNSE-en and CNSS-en.The experimental results show that the F1 values of the HNNR method are 11%,3%,and 2% higher than baseline matching models Match Pyramid,LDA,and CIG-Sim-GCN,respectively.