Keep it simple: random oversampling for imbalanced data

The issue of imbalanced data affects a wide range of applications. Despite a plethora of sophisticated sampling techniques for dealing with imbalanced data, the simple random oversampling (ROS) method remains a robust alternative. The goal of this paper is to compare the performance of ROS to the more advanced sampling algorithms. To this end, we conduct numerical experiments on multi-label data. The results of the experiments reveal that ROS outperforms several advanced sampling algorithms. Given the computational efficiency of ROS and its robust accuracy, we believe that it provides a good option for dealing with imbalanced data.