With the rapid development of information age, recommendation systems have played an important role in many fields. However, traditional recommendation systems still face challenges in information overload, cold-start problem and inaccurate recommendations. To address the issue of inaccurate recommendations, this paper presents a scholar recommendation system based on the think tank, an academic information retrieval platform. First, we use the SCHOLAT data to establish the think tank social network, and established the think tank scholar word cloud by using the HanLP model and the TF-IDF algorithm. Based on this, we constructed a think tank dataset. Subsequently, we use a modified Variational Graph Normalized Auto-Encoders(VGNAE) model to perform the scholar recommendation task on the think tank dataset and propose an efficient scholar update method. Finally, we deposit the recommendation results in the cache to improve the recommendation efficiency. The experimental results show that compared with other baseline models, the improved VGNAE model performs well on the think tank dataset, and achieves the best results on both AUC and AP, verifying the effectiveness and reliability of our proposed think tank recommendation system.