Improving Detection of Rare Nodes in Hierarchical Multi-Label Learning

In hierarchical multi-label classification, a persistent challenge is enabling model predictions to reach deeper levels of the hierarchy for more detailed or fine-grained classifications. This difficulty partly arises from the natural rarity of certain classes (or hierarchical nodes) and the hierarchical constraint that ensures child nodes are almost always less frequent than their parents. To address this, we propose a weighted loss objective for neural networks that combines node-wise imbalance weighting with focal weighting components, the latter leveraging modern quantification of ensemble uncertainties. By emphasizing rare nodes rather than rare observations (data points), and focusing on uncertain nodes for each model output distribution during training, we observe improvements in recall by up to a factor of five on benchmark datasets, along with statistically significant gains in $F_{1}$ score. We also show our approach aids convolutional networks on challenging tasks, as in situations with suboptimal encoders or limited data.

💡 Research Summary

Hierarchical multi‑label learning (HML) poses a unique challenge: the hierarchical constraint forces parent‑node probabilities to dominate those of their children, which naturally creates a long‑tailed distribution where deep, fine‑grained nodes are extremely rare. Existing remedies—such as resampling, cost‑sensitive learning, or observation‑level weighting—tend to over‑emphasize common parent nodes because they are present in every sample that contains a rare child. Consequently, rare node detection remains poor, especially in full‑depth (HML‑FD) settings where annotations reach the leaf level.

The authors address this problem by reframing imbalance from an observation perspective to a node‑centric perspective and by augmenting the hierarchical loss with two complementary weighting schemes: (1) a node‑wise imbalance weight derived from each node’s frequency in the training set, and (2) a focal‑style weight that leverages modern ensemble‑based uncertainty estimates to focus learning on nodes where the model is uncertain.

Imbalance weighting
For each node i they compute a raw weight wᵢ = (N_obs·N_classes / nᵢ) – 1, where nᵢ is the total count of occurrences of node i and all its descendants. To avoid extreme values that could destabilize training, they rescale wᵢ to a