Abstract:Imbalanced dataset tends to be biased towards "majority" when classifying, and samples generated by traditional over-sampling cannot well express the distribution characteristics of the original dataset. The improved variational autoencoders combine with data preprocessing method, generate samples by the generator of variational autoencoders trained by the minority class samples to balance the training data set, solve the overfitting problem caused by imbalanced dataset of traditional sampling. Experiments are carried out on four commonly used UCI datasets, the results demonstrate that the proposed method shows better classification performance in F_measure and G_mean with high accuracy.