Binary relevance (BR) is regarded as the most intuitive solution to learn from multi-label data, which decomposes the multi-label learning task into a number of independent binary learning tasks (one per class label). To amend its potential weakness of ignoring label correlations, many correlation-enabling extensions to BR have been proposed based on two major strategies, i.e. assuming random correlations with chaining structure or taking full-order correlations with stacking structure. However, in both strategies label correlations are only exploited in an uncontrolled manner, which may be problematic when error-prone and uncorrelated class labels arise. In this paper, to fulfill controlled label correlations exploitation, a novel enhancement to BR is proposed based on a two-stage filtering procedure. In the first stage, error-prone class labels are pruned from the label space based on holdout validation. In the second stage, closely-related class labels are identified based on supervised feature selection by viewing unpruned labels as features. Extensive experiments across fourteen multi-label data sets confirm the superiority of controlled label correlations exploitation, especially when large number class labels exist in the label space.
展开▼