Multi-label classification faces a fundamental tension: modeling complex label interactions effectively often requires large, computationally intensive architectures, while simpler models fail to capture crucial dependencies. This work introduces Deep Extra-Trees (DET) , a layered, hierarchical ensemble based on Extremely Randomized Trees that refines predictions by augmenting original features with probabilistic context from previous layers. In DET, the term “deep” refers strictly to multi-layer stacking of ensembles (not neural representation learning or backpropagation), enabling progressive refinement of feature–label interactions through iterative prediction augmentation. DET provides a unifying framework for major multi-label transformation strategies—Binary Relevance, Classifier Chains, and Label Powerset—so a single architecture can adapt to diverse correlation structures. Experiments on eight benchmark datasets show that DET, particularly its Classifier Chains variant, improves Macro-F1 over strong baselines including ECC and RAkEL, while retaining the interpretability and efficiency of tree ensembles. These results indicate that hierarchical refinement can be achieved within a transparent ensemble framework, providing a practical and scalable alternative to purely neural approaches.
Berrouachedi et al. (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: