Combining multi-label classifiers based on projections of the output space using Evolutionary algorithms

Abstract

The multi-label classification task has gained a lot of attention in the last decade thanks to its good application to many real-world problems where each object could be attached to several labels simultaneously. Several approaches based on ensembles for multi-label classification have been proposed in the literature; however, the vast majority are based on randomly selecting the different aspects that make the ensemble diverse and they do not consider the characteristics of the data to build it. In this paper we propose an evolutionary method called Evolutionary AlGorithm for multi-Label Ensemble opTimization, EAGLET, for the selection of simple, accurate and diverse multi-label classifiers to build an ensemble considering the characteristics of the data, such as the relationship among labels and the imbalance degree of the labels. In order to model the relationships among labels, each classifier of the ensemble is focused on a small subset of the label space, resulting in models with a relative low computational complexity and lower imbalance in the output space. The resulting ensemble is generated incrementally given the population of multi-label classifiers, so the member that best fits to the ensemble generated so far, considering both predictive performance and diversity, is selected. The experimental study comparing EAGLET with state-of-the-art methods in multi-label classification over a wide set of sixteen datasets and five evaluation measures, demonstrated that EAGLET significantly outperformed standard MLC methods and obtained better and more consistent results than state-of-the-art multi-label ensembles.