With the development of personalized book recommendation systems, accurate user reading behavior portrait construction has become the key to improving recommendation effects. Traditional user portrait construction is based on static data features, which makes it difficult to capture the temporal changes in user interests and has poor expression of deep semantic associations between articles. This article uses the knowledge graph embedding method of TA and TransD models to construct user portraits and improve the quality of constructed user portraits. The study first captures entities based on user reading behavior, constructs a topological graph, and introduces the TransD knowledge graph embedding method to enhance the user portrait’s ability to understand the deep semantics of the article content. Then the TA mechanism is used to model the user’s reading behavior and capture the dynamic changes of user interests. Finally, the temporal interest modeling is combined with the knowledge graph embedding to construct a more accurate user portrait in a weighted fusion manner. The results show that the TransD-TA model has the highest number of overlaps in user attribute matching, which is 179, an increase of 34 compared to the traditional K-means, and the construction time is only 28 ms when the number of users is 100. The experimental results show that the TransD-TA method can effectively improve the accuracy of user portraits and promote their performance in personalized recommendation tasks.