<p>Imbalanced data is a common characteristic in classification problems. The imbalance of the data may lead to biased results, due to the difficulty of learning classes with few observations, called Minority classes. Balancing techniques are a common strategy to overcome imbalanced data situations. Real problems may have atypical observations or outliers, which can have a negative impact on the characterization of the Minority classes. Upon application of a balancing technique, the effect of these atypical observations may still be present in the balanced dataset, and their negative effects may even be reinforced. In this paper we propose a robust approach to imbalanced learning—which we call RM-SMOTE—that combines the idea of SMOTE with robust Mahalanobis distance. We propose to down-weight atypical Minority class observations in such a way that potential outliers from this class have a low chance of being selected in the resampling step. The performance of the RM-SMOTE is evaluated using several simulation scenarios, with different levels of contamination and imbalance ratios, and also using a set of 24 imbalanced benchmark datasets. For each dataset, we consider other balancing techniques, and we provide several performance measures, allowing for comparison between the techniques. We also assess the impact in the classification results of the proportion of outliers in the data. Statistical tests support the claim that the RM-SMOTE is among the best balancing techniques. In the few cases where RM-SMOTE does not outperform other balancing techniques for some performance metrics related with the Minority class, the results show that it has an overall better performance in the sense that our method improves the results for the Minority class without penalizing too much the performance related with the Majority class. The implementation of RM-SMOTE is available at <a href="https://github.com/rasooltaban/RM-SMOTE">https://github.com/rasooltaban/RM-SMOTE</a>.</p>