Evolutionary multiobjective optimization for multilabel learning

Abstract

Machine learning is traditionally considered as a single objective optimization problem. That is, a learner iteratively improves performances through optimizing one single objective function. However, many machine-learning tasks need to simultaneously optimize multiple objective functions. For example, ensemble learning need to find a set of accurate but diverse base learners, and it is desirable for the feature selection task to extract a small number of representative features. As a consequence, there is a surge of multiobjective machine learning in recent years. That is, a machine-learning task is considered as the multiobjective optimization problem (MOP). Moreover, evolutionary algorithms are usually employed to optimize it, since evolutionary algorithms have shown their superiority for MOPs. In this chapter, we will introduce our recent work in multilabel learning with multiobjective optimization. As an important supervised learning task, multilabel learning refers to the task of predicting potentially multiple labels for a given instance. Conventional multilabel learning approaches focus on a single objective setting, and there is a basic assumption that the optimization over one single objective can improve the overall performance of multilabel learning. However, in many real applications, an optimal multilabel learner may need to consider the trade-offs among multiple objectives. In this chapter, we will present two works in evolutionary multiobjective optimization for multilabel learning. In the first work, we directly optimize multiobjective functions based the multiobjective optimization framework; in the second work, we generate a set of accurate but diverse multilabel learners with the ensemble learning framework.