MLSMOTE: Approaching imbalanced multilabel learning through synthetic instance generation

Learning from imbalanced data is a problem which arises in many real-world scenarios, so does the need to build classifiers able to predict more than one class label simultaneously (multilabel classification). Dealing with imbalance by means of resampling methods is an approach that has been deeply studied lately, primarily in the context of traditional (non-multilabel) classification. In this paper the process of synthetic instance generation for multilabel datasets (MLDs) is studied and MLSMOTE (Multilabel Synthetic Minority Over-sampling Technique), a new algorithm aimed to produce synthetic instances for imbalanced MLDs, is proposed. An extensive review on how imbalance in the mul-tilabel context has been tackled in the past is provided, along with a thorough experimental study aimed to verify the benefits of the proposed algorithm. Several multilabel classification algorithms and other multilabel oversampling methods are considered, as well as ensemble-based algorithms for imbalanced multilabel classification. The empirical analysis shows that MLSMOTE is able to improve the classification results produced by existent proposals. Classification is one of the main supervised learning applications , an important field in Machine Learning [1]. The goal is to train a model using a set of labeled data samples, obtaining a clas-sifier able to label new, never seen before, unlabeled samples. The datasets used in traditional classification have only one class per instance. By contrast, in multilabel datasets (MLDs) [2] each instance has more than one class assigned, and the total number of different classes (labels) can be huge. In many real world scenarios, such as text classification [3] and fraud detection [4], the number of instances associated to some classes is much smaller (greater) than the amount of instances assigned to others. This problem, known as imbalanced learning, has been widely studied over the last decade [5] in the context of classic classification. It is also present in multilabel classification (MLC), since labels are unevenly distributed in most MLDs. To deal with imbalance in MLC, methods based on algorithmic adaptations [6–8], the use of ensembles [9,10], and resampling techniques [11– 13] have been proposed. Among the existent resampling techniques, those based on the generation of new samples (oversampling) have shown [14] to work better than others. The new samples can be clones of existent ones, or be synthetically produced as in SMOTE (Synthetic Minority Over-sampling Technique) [15]. Multilabel oversampling algorithms based on the cloning approach have been proposed in [12,13], being demonstrated its capability to deliver …