Recommender systems often exhibit a significant popularity bias such as the collected interaction data typically shows quite imbalanced or even long-tailed distribution over items. This distribution may result from user conformity to the group, which fails to reflect their true preferences. Existing methods for alleviating popularity bias usually adjust the regularization in the loss function or indirectly modify the loss function via causal graphs. However, these methods suffer from a cascading effect during the aggregation process of Graph Neural Networks (GNNs) and it is difficult to mitigate both data and model biases. To this end, we propose MLDRF, a Multi-Level Debiasing Recommendation Framework. First, we introduce a neighbor reweighting mechanism to mitigate popularity bias in GNN aggregation, thereby effectively reducing data bias. We then use causal graphs to disentangle the popularity bias, enhancing the ability of the GNN to handle complex biases. Finally, Zerosum regularization is applied to optimize the loss function, thus reducing the model bias. Comprehensive experiments on four public datasets show an improvement of 3% to 5% in recommendation accuracy and the overall performance is improved compared to the state-of-the-art methods.