| News recommendation plays an essential role in the field of recommendation,with the primary purpose of providing users with news information that they are interested in by mining their reading interests.These news articles usually contain rich textual information,such as titles,summaries,categories,and body text.In recent years,news-based recommendation methods have developed rapidly,which generate news representations by fusing features from different fields of news information.However,these news modeling methods still need to improve,such as insufficient information extraction and a lack of semantic connections between different information fields.Feature extraction for news recommendation mainly considers the perspectives of news and users to provide users with news content they are interested in.In response to the above issues,this article proposes a news recommendation method called NMSF(News Recommendation With Multi-task And Semantic-level Cross-domain Feature Fusion).This method performs a cross-domain fusion of different domain information of news at the semantic level.It introduces a multi-task structure to improve the performance of the click rate prediction module.Specifically,in the news representation generation stage,NMSF first extracts features of news titles,summaries,and categories and concatenates them at the semantic level to generate news representations.In the user representation generation stage,a multi-head self-attention mechanism is used to weigh and aggregate the user’s historical news representation sequence to generate the user’s interest representation.In the click prediction stage,we divide user representations into user content representations and user category representations and predict with candidate news content and candidate news categories,respectively.Finally,the predicted results of these two tasks are weighted and summed to obtain the final click probability for the candidate news.This article also proposes a new news recommendation method called NWFF(News Recommendation With Word-level Cross-domain Feature Fusion),which performs a cross-domain fusion of different domain information of news at the word level.In the generation stage of news representation,NWFF combines the category,title,and summary of news to generate cross-domain information pairs.Then,these cross-domain matching representations are aggregated into news representations through a multi-head self-attention mechanism.This article verifies the superiority of NMSF and NWFF methods on Microsoft’s MIND dataset.The experimental results show that NMSF and NWFF significantly improved AUC,MRR,NDCG@5,and NDCG@10 indicators compared with the comparison models.The NWFF and NMSF methods proposed in this article address the need for cross-domain information fusion and insufficient information in news recommendations.In addition,these methods have also made innovative improvements in cross-domain information fusion and multi-task applications,providing effective research methods for subsequent researchers. |
