The paper introduces a Double Correction Framework for Denoising Recommendation (DCF) to address the issue of noisy samples in implicit feedback data, which is common in recommender systems. Traditional methods often drop noisy samples during training, but this approach has limitations, such as ignoring the instability of optimization and the sparsity of data. To overcome these issues, DCF proposes two main components: sample dropping correction and progressive label correction. 1. **Sample Dropping Correction**: This component aims to stabilize the model's predictions by calculating the mean loss over time intervals and using a damping function to reduce the impact of outliers. It also identifies and retains hard samples, which are informative and beneficial for recommendation performance, by deriving a lower bound for the loss using concentration inequalities. 2. **Progressive Label Correction**: This component iteratively re-labels highly deterministic noisy samples and re-trains them to improve performance. The relabeling strategy is adapted to the model's optimization process, gradually increasing the proportion of relabeled samples as training progresses. The paper demonstrates the effectiveness and generalization of DCF through extensive experiments on three datasets and four backbones. The results show that DCF outperforms existing methods in terms of NDCG@K and Recall@K metrics, confirming the benefits of the proposed framework in mitigating the adverse effects of noisy samples. The code for DCF is available at https://github.com/bruno686/DCF.The paper introduces a Double Correction Framework for Denoising Recommendation (DCF) to address the issue of noisy samples in implicit feedback data, which is common in recommender systems. Traditional methods often drop noisy samples during training, but this approach has limitations, such as ignoring the instability of optimization and the sparsity of data. To overcome these issues, DCF proposes two main components: sample dropping correction and progressive label correction. 1. **Sample Dropping Correction**: This component aims to stabilize the model's predictions by calculating the mean loss over time intervals and using a damping function to reduce the impact of outliers. It also identifies and retains hard samples, which are informative and beneficial for recommendation performance, by deriving a lower bound for the loss using concentration inequalities. 2. **Progressive Label Correction**: This component iteratively re-labels highly deterministic noisy samples and re-trains them to improve performance. The relabeling strategy is adapted to the model's optimization process, gradually increasing the proportion of relabeled samples as training progresses. The paper demonstrates the effectiveness and generalization of DCF through extensive experiments on three datasets and four backbones. The results show that DCF outperforms existing methods in terms of NDCG@K and Recall@K metrics, confirming the benefits of the proposed framework in mitigating the adverse effects of noisy samples. The code for DCF is available at https://github.com/bruno686/DCF.