| Tens of thousands of news stories are produced daily on various news websites due to the growth of the Internet and integrated media,making it more challenging for readers to find the news they are interested in among the vast volume of data.The news recommendation system was created in order to address the issues of efficacy,timeliness,and preference when users browse in the face of enormous amounts of news data.The core of a news recommendation system is its algorithm,and whether it is good or bad will directly affect the effectiveness of its recommendations.However,there are still many issues with the news recommendation algorithm that need to be resolved,such as the fact that the characteristics that make news different from other texts are not taken into account when modeling news representation;the change of users’ reading interests is variable,and in the news sequence of the same user’s reading history,the user’s interest in different news is different,and how to accurately model the user’s interest from the user’s reading history.Furthermore,present news recommendation algorithms tend to neglect the novelty of the user’s suggestion results,lack user preference exploration,and are incapable of adapting to the quick change of user interests.To address the above issues,this work conducts research on a personalized news recommendation approach based on representation learning,and the primary work is as follows.1)A news encoder that integrates entity features is presented to solve the problem that existing news recommendation algorithms model news texts without taking into account the properties of news texts that are independent of other texts,and the modeled news vectors frequently fail to effectively represent news.The news encoder is based on the principle of representation learning,which extracts named entities in the text as a high level summary of news content and uses multiple attention mechanisms to extract news features and obtain a high-quality news vector representation.2)A user interest modeling encoder based on user reading history is suggested as a solution to the issue of challenging user interest modeling.In order to express a user’s preference for different news to a neural network,the approach extracts the attention of various news from the user’s recently read news sequence.Specifically,based on the representation learning method,multiple attention mechanisms are used to model the user’s recent news reading sequence to extract the user’s attention to different news.In order to further engage with the candidate news and obtain the ultimate user interest representation,an activation unit based on the Scaled Dot-Product attention mechanism is also developed.3)A personalized news recommendation model based on knowledge graph(KGMA)is proposed in order to satisfy the users’ demand for novelty in pushing news.KGMA enhances news features by introducing knowledge graph information and enriches contextual information by utilizing the Trans E model for knowledge graph embedding of news knowledge graph subgraphs. |
