A diversity and reliability-enhanced synthetic minority oversampling technique for multi-label learning

The class imbalance issue is generally intrinsic in multi-label datasets due to the fact that they have a large number of labels and each sample is associated with only a few of them. This causes the trained multi-label classifier to be biased towards the majority labels. Multi-label oversampling methods have been proposed to handle this issue, and they fall into clone-based and Synthetic Minority Oversampling TEchnique-based (SMOTE-based) ones. However, the former duplicates minority samples and may result in over-fitting whereas the latter may generate unreliable synthetic samples. In this work, we propose a Diversity and Reliability-enhanced SMOTE for multi-label learning (DR-SMOTE). In it, the minority classes are determined according to their label imbalance ratios. A reliable minority sample is used as a seed to generate a synthetic one while a reference sample is selected for it to confine the synthesis region. Features of the synthetic samples are determined probabilistically in this region and their labels are set identically to those of the seeds. We carry out experiments with eleven multi-label datasets to compare DR-SMOTE against seven existing resampling methods based on four base multi-label classifiers. The experimental results demonstrate DR-SMOTE's superiority over its peers in terms of several evaluation metrics.