In recent years, the point-of-interest (POI) recommendation system has gradually become one of the research hotspots in the field of mobile recommendation systems. The method of joint modeling of multiple factors, such as time, space, sequence, socialization, and semantic information, has been gradually introduced into a unified model to compute the user preferences under multidimensional scenarios. As an effective multi-factor joint modeling method, the embedding learning model has better performance in the mobile recommendation systems. However, many of the embedded learning models just simply embed the explicit factors, such as timestamps, items, regions, sequences, etc. into the same space. Due to the lack of deep mining of user and item semantic features, it is hard to accurately obtain user preferences when the users’ check-in data is extremely sparse. In view of this, a multi-dimensional context-aware graph embedding model, called MCAGE, is proposed in this study. In MACGE model, the topic model is used to extract the potential semantic features between users and items. Then, a series of graph nodes and association rules are redefined. To enhance the accuracy of describing the user preferences, a more effective user preference formula is designed. Finally, the results of experiments based on the real-world dataset shows that the proposed model has better recommendation performance.