Dynamic ensemble learning for multi-label classification

Abstract

Ensemble learning has been shown to be an effective approach to solve multi-label classification problem. However, most existing ensemble learning methods do not consider the difference between unseen instances, and existing methods that consider such difference fail to effectively explore label correlation, which limits their performance. To address these issues, we propose a novel method named MLDE (Multi-Label classification with Dynamic Ensemble learning). In MLDE, the most competent ensemble of base classifiers is selected and combined to predict each unseen instance. To make dynamic selection specific to multi-label problem and achieve better performance, we combine classification accuracy and ranking loss to serve as the competence measurement for the base classifiers. Specifically, classification accuracy is decomposable to multiple labels and distinguishes the ability difference of a classifier with respect to different labels, while ranking loss focuses on the overall performance of a classifier on the label set and thus fully considers the correlation between multiple labels. Extensive experiments on 24 publicly available datasets demonstrate that MLDE outperforms the state-of-the-art methods.