Correlation-based pruning of dependent binary relevance models for Multi-label classification

Binary relevance (BR), a basic Multi-label classification (MLC) method, learns a single binary model for each different label without considering the dependences among rest of labels. Many chaining and stacking techniques exploit the dependences among labels to improve the predictive accuracy for MLC. Using these two techniques, BR has been promoted as dependent binary relevance (DBR). In this paper we propose a pruning method for DBR, in which the Phi coefficient function has been employed to estimate correlation degrees among labels for removing irrelevant labels. We conducted our pruning algorithm on benchmark multi-label datasets, and the experimental results show that our pruning approach can reduce the computational cost of DBR and improve the predictive performance generally.