As internet technologies continue to advance, users are increasingly overwhelmed by excessive information, making it crucial to have effective systems for personalized content suggestions. However, when new users or insufficient data are present, traditional recommendation systems can suffer from issues like limited data and the cold start problem, which can degrade the quality of recommendations. In this research, a new movie knowledge graph model is proposed using the Movielens-1M dataset, alongside an adaptive recommendation method that combines graph-based learning and attention mechanisms. The approach improves the performance and interpretability of the knowledge graph by refining entity similarity calculations and optimizing the simplification of connection paths. It also strengthens the recommendation system’s capacity to perform well even with minimal data. Results from the experiments indicate that the proposed method excels in both prediction accuracy and user satisfaction, particularly in overcoming challenges related to data insufficiency and initial user interactions.