Learning Gradient Boosted Multi-label Classification Rules

In multi-label classification, where the evaluation of predictions is less straightforward than in single-label various meaningful, though different, loss functions have been proposed. Ideally, learning algorithm should be customizable towards a specific choice performance measure. Modern implementations boosting, most prominently gradient boosted decision trees, appear to appealing from this point view. However, they are mostly limited and hence not amenable losses unless these label-wise decomposable. work, we develop generalization boosting framework multi-output problems propose an for classification rules that able minimize decomposable as well non-decomposable functions. Using well-known Hamming subset 0/1 representatives, analyze abilities limitations our approach on synthetic data evaluate its predictive benchmarks.