A new approximate query processing method was proposed to solve the problem of low query accuracy caused by data skew in the approximate query processing. First of all, the algorithm was based on the conditional generative adversarial network and incorporated the conditional variational auto-encoder to ensure the stability and the accuracy of the algorithm. The Wasserstein distance was used to measure the model error to eliminate model collapse. Secondly, based on the above generative model, approximate query processing was achieved and users’ queries were answered without accessing the underlying data, avoiding disk interaction. The model was combined with aggregate precomputation to form an efficient approximate query processing framework to answer interactive queries more accurately and quickly. Finally, an efficient voting algorithm was designed to filter the samples generated by the model and the internal data of the samples, so as to improve the quality of the generated samples and minimize the query error. Experimental results show that, compared with other approximate query processing algorithms, the method proposed can effectively overcome the influence of data skew and answer queries more accurately in shorter interaction time.