<p>Missing data seriously threatens data quality as it leads to the loss of important information. Therefore, addressing missing data is critical for ensuring accurate and reliable results in data mining. Based on generative adversarial networks (GAN), this paper proposes a novel data imputation method named dynamic multiple generative adversarial networks (DMGAN). DMGAN comprises three major components: the improved missing weight KNN (MWKNN) algorithm, the attribute imputation priority, and the nearest neighbor-based generative adversarial imputation networks (NNGAIN). In DMGAN, the incomplete dataset is initially pre-imputed using MWKNN to utilize the information in incomplete samples. Then, the dataset is partitioned into multiple parts based on data label and attribute imputation priority, and each part is dynamically imputed in order using NNGAIN. During this process, the imputed parts are dynamically put into the training of subsequent models to leverage the value of the obtained imputation results. Additionally, the nearest neighbor information and imputation priority are updated simultaneously based on the latest results of MWKNN and NNGAIN. DMGAN takes into account both sample distribution and attribute characteristics so that the quality of imputed results is enhanced. To verify the proposed method’s effectiveness, DMGAN and nine other popular comparison methods, including state-of-the-art GAN-based imputation techniques, are implemented to impute seven public datasets and two clinically obtained Alzheimer’s disease datasets. The imputed results of DMGAN achieve superior RMSE performance compared to other algorithms, surpassing the second-best method by up to 33.6%. In classification tasks, DMGAN’s imputed results also demonstrate significant improvements, with F1 score increasing by up to 33.0% and AUC by up to 15.6% over other methods. The experimental results indicate that DMGAN generally provides higher-quality imputed results in various cases. </p>