Crash data augmentation using variational autoencoder

•In this paper, we present a data augmentation technique to reproduce crash data.•Variational Autoencoder (VAE) was used to generate millions of crash samples from only a limited number of training data.•The generated data was compared to real data from different statistical standpoints and similarity was reported.•It was also compared to some of the minority oversampling techniques like SMOTE and ADASYN. The results were also compared with the GAN framework for generating data.•Crash prediction models based on Logistic Regression, Support Vector Machine and Artificial Neural Network were used to compare the generated data from the different models.•Overall, VAE showed excellent results compared to the other data augmentation methods. In this paper, we present a data augmentation technique to reproduce crash data. The dataset comprising crash and non-crash events are extremely imbalanced. For instance, the dataset used in this paper consists of only 625 crash events for over 6.5 million non-crash events. Thus, learning algorithms tend to perform poorly on these datasets. We have used variational autoencoder to encode all the events into a latent space. After training, the model could successfully separate crash and non-crash events. To generate data, we sampled from the latent space containing crash data. The generated data was compared with the real data from different statistical aspects. t-Test, Levene-test and Kolmogrove Smirnov test showed that the generated data was statistically similar to the real data. It was also compared to some of the minority oversampling techniques like SMOTE and ADASYN as well as the GAN framework for generating data. Crash prediction models based on Logistic Regression (LR), Support Vector Machine (SVM) and Artificial Neural Network (ANN) were used to compare the generated data from the different oversampling techniques. Overall, variational autoencoder (VAE) showed excellent results compared to the other data augmentation methods. Specificity is improved by 8% and 4% for VAE-LR and VAE-SVM respectively when compared to SMOTE while the sensitivity is improved by 6% and 5% when compared to ADASYN. Moreover, VAE generated data also helps to overcome the overfitting problem in SMOTE and ADASYN since there is flexibility in choosing the decision boundary.