Improving multilabel classification via heterogeneous ensemble methods

Abstract

We consider the task of multilabel classification, where each instance may belong to multiple labels simultaneously. We propose a new method, called multilabel super learner (MLSL), that is built upon the problem transformation approach using the one-vs-all binary relevance method. MLSL is an ensemble model that predicts multilabel responses by integrating the strength of multiple base classifiers, and therefore it is likely to outperform each base learner. The weights in the ensemble classifier are determined by optimization of a loss function via $V$-fold cross-validation. Several loss functions are considered and evaluated numerically. The performance of various realizations of MLSL is compared to existing problem transformation algorithms using three real data sets, spanning applications in biology, music, and image labeling. The numerical results suggest that MLSL outperforms existing methods most of the time evaluated by the commonly used performance metrics in multilabel classification.