Knowledge-driven Personalized News Recommendation

Online news service platforms such as Toutiao and Microsoft News have become important platforms for users to obtain news information.Massive news articles are generated and published online every day,making it difficult for users to find interested news quickly.However,personalized news recommendation can help users alleviate information overload and improve news reading experience.Thus,it is widely used in many online news platforms.There are two important problems in news recommendation,i.e.,how to represent news articles which have rich textual content and how to model users' interest in news from their previous behaviors.For the representation of news texts,in this thesis we uses a modified transformer model and integrates news category information to learn the news context representation,because the multi-head self-attention network can model the interaction between words,and news category information can also reflect user preferences.As for user interest modeling,the following two aspects are studied in this thesis:(1)A hierarchical hash network method is proposed to enhance the user's interest representation.For users with few news browsing records,it is also called the user cold start problem.While most existing news recommendation methods usually learn user representations from interactive historical news,and cannot model user interests well when users are cold-started.Therefore,this thesis considers the clustering characteristics among users,a hierarchical hash network is used to map low-informative users to rich-informative frequent users to obtain interest representation of the hash part.At the same time,the historical part of the interest representation is still obtained from the news reading records,and finally the hash vector and the historical vector are combined through the route attention mechanism to generate the final user interest representation.(2)A knowledge-guided reinforcement learning model is proposed to dynamically capture user interest changes.For a user with a rich news browsing history,the users' interests are likely to be complicated and time-varying.At present,many deep learning methods only focus on the local benefits of the recommended news,they are more fitting to the existing historical data of clicked news,which makes it easier to recommend homogeneous content to users.Therefore,this thesis adopts the reinforcement learning method to consider the long-term impact of a certain news clicked by the user on the sequence,and captures the change of interest in real time by obtaining feedback after interacting with the user's click action.At the same time,knowledge graph information is integrated to guide the exploration process of reinforcement learning effectively,so that the reinforcement learning model can well capture user preference drift in the recommendation task.In this thesis,the hierarchical hash network method and the knowledge guided reinforcement learning model are compared with a variety of news recommendation methods,and a large number of comparative experiments are carried out on Microsoft News dataset(MIND).They are better than the baseline model in the case of user cold start and user with rich clicked records,which verifies the effectiveness of the news recommendation method proposed in this thesis.