With the pervasiveness of smart equipment and social networking, it has become an urgent task to recommend personalized news to users. For news recommendation, researchers have proposed to finish the news recommendation task based on collaborative filtering algorithms, and also proposed to employ additional information solving high data sparsity problem. A typical method is to use the structure of knowledge graph to represent heterogeneous knowledge. But in existing methods, it is difficult to define the relationship among entities, and can only use the static knowledge without utilization and consideration of the evolution of knowledge graph. To solve these problems, in this paper, we first propose a novel algorithm to efficiently extract the most relevant tags in the news to represent the news. Then we construct the structure of the knowledge graph to represent the personal information and historical click records of a user. Those extracted knowledge is used to build a user profile and further to form a community of users, which can also alleviate the cold start problem. We conducted experiments on a real-world news website, including the labels extraction. We report the experimental results of metrics evaluation and manual evaluation, and the results demonstrate the effectiveness and efficiency of the proposed framework.