Multilabel Classification with Partial Abstention: Bayes-Optimal Prediction under Label Independence