Recently, recommender system has been widely used on electronic commerce websites and other fields. The graph convolutional neural network (GCN), an emerging deep learning model, is used to excavate network topology information to improve the prediction precision for recommender systems. Previous research has shown that deeper models, such as GCN, can extract more enriched features. However, researchers have proposed that the deeper model is not always better due to the disappearance of the gradient, caused by network weights’ ineffective update. In this paper, a hybrid deep learning model called GraphRGr is proposed to obtain desired recommendation result with residual network (ResNet) and GCN for recommendations to solve degradation problems. In order to ensure the robustness of GraphRGr against shilling attacks, an end-to-end learning process named neural random forests (NRF) is used to resist shilling attacks, which makes GraphRGr have an excellent recommendation performance than other models whether there exists shilling attacks or not. Extensive experiments demonstrated the superior performance of the proposed GraphRGr over the baseline models.