Data augmentation is one of the most successful techniques to improve the classification accuracy of machine learning models in computer vision. However, applying data augmentation to tabular data is a challenging problem since it is hard to generate synthetic samples with labels. In this paper, we propose an efficient classifier with a novel data augmentation technique for tabular data. Our method called CCRAL combines causal reasoning to learn counterfactual samples for the original training samples and active learning to select useful counterfactual samples based on a region of uncertainty. By doing this, our method can maximize our model's generalization on the unseen testing data. We validate our method analytically, and compare with the standard baselines. Our experimental results highlight that CCRAL achieves significantly better performance than those of the baselines across several real-world tabular datasets in terms of accuracy and AUC. Data and source code are available at: https://github.com/nphdang/CCRAL.