An evolutionary approach to build ensembles of multi-label classifiers

Abstract

In recent years, the multi-label classification task has gained the attention of the scientific community given its ability to solve problems where each of the instances of the dataset may be associated with several class labels at the same time instead of just one. The main problems to deal with in multi-label classification are the imbalance, the relationships among the labels, and the high complexity of the output space. A large number of methods for multi-label classification has been proposed, but although they aimed to deal with one or many of these problems, most of them did not take into account these characteristics of the data in their building phase. In this paper we present an evolutionary algorithm for automatic generation of ensembles of multi-label classifiers by tackling the three previously mentioned problems, called Evolutionary Multi-label Ensemble (EME). Each multi-label classifier is focused on a small subset of the labels, still considering the relationships among them but avoiding the high complexity of the output space. Further, the algorithm automatically designs the ensemble evaluating both its predictive performance and the number of times that each label appears in the ensemble, so that in imbalanced datasets infrequent labels are not ignored. For this purpose, we also proposed a novel mutation operator that considers the relationship among labels, looking for individuals where the labels are more related. EME was compared to other state-of-the-art algorithms for multi-label classification over a set of fourteen multi-label datasets and using five evaluation measures. The experimental study was carried out in two parts, first comparing EME to classic multi-label classification methods, and second comparing EME to other ensemble-based methods in multi-label classification. EME performed significantly better than the rest of classic methods in three out of five evaluation measures. On the other hand, EME performed the best in one measure in the second experiment and it was the only one that did not perform significantly worse than the control algorithm in any measure. These results showed that EME achieved a better and more consistent performance than the rest of the state-of-the-art methods in MLC.