Abstract: Existing deep learning-based recommendation models have mainly focused on improving the accuracy of recommendation systems. However, beyond recommendation accuracy, the reliability of the model's recommendations is also of great concern. Therefore, a rating-trustworthy recommendation model based on generative adversarial networks (GANs) is proposed to evaluate the effectiveness of prediction results and achieve a balance between recommendation accuracy and reliability. This model solely employs explicit user rating information to gauge the credibility of predicted ratings and screens out highly credible predicted ratings based on a predefined reliability threshold, thus ensuring the trustworthiness of recommended items. Furthermore, to enhance the prediction performance of the model and ensure fairness in training, a positive sample padding strategy is designed to mitigate the data imbalance problem in the rating reliability matrix. Experimental results on three real datasets show that the proposed model outperforms selected comparison methods in both Recall and NDCG metrics, effectively improving the performance of recommendation systems.