Hierarchical Labeling and Supervision

• Hierarchical labeling and supervision are techniques that use structured taxonomies (trees or DAGs) to incorporate semantic hierarchies into learning systems.
• They improve model performance by embedding hierarchy-aware losses, multi-path supervision, and graph-based label modeling to reduce gross misclassifications.
• These methods are applied across fully and weakly supervised scenarios, enabling robust, continual learning and fine-grained discovery with minimal annotation.

Hierarchical labeling and supervision refer to the use of explicit, structured label taxonomies—ranging from simple trees to general directed acyclic graphs (DAGs)–in the design, supervision, and evaluation of learning systems. These methods exploit the fact that semantic categories naturally form hierarchies (e.g., species → genus → family, or object → part → subpart) to enhance model accuracy, generalization, and efficiency, especially in regimes where labels are scarce, ambiguous, or costly to obtain. Hierarchical supervision can be deployed in fully supervised, semi-supervised, and open-set or continual learning scenarios, and underlies several recent advances at the interface of deep representation learning, contrastive methods, and probabilistic modeling.

1. Hierarchical Label Structures and Taxonomies

Label hierarchies are formally encoded as trees or DAGs where each node corresponds to a label (possibly internal or a leaf), and the edge relations encode parent–child (hypernym–hyponym) inclusion (Urbani et al., 2024, Stoimchev et al., 12 Mar 2026, Zhang et al., 2024). In many real-world datasets (ImageNet, iNaturalist, DBPedia, functional genomics), labels are distributed over multi-level structures, with leaves representing fine-grained categories and internal nodes corresponding to superclasses.

A key property is path consistency: when an instance is assigned a label at any depth, it must also be assigned all its ancestor labels. Multi-label and multi-path settings allow a single instance to activate multiple independent label paths, reflecting compositional structure (e.g., a scene with "water" and "urban" tags) (Stoimchev et al., 12 Mar 2026, Yu et al., 2023).

Hierarchical labels support operations such as:

• Inducing coarser/finer partitions,
• Measuring semantic distances (e.g., LCA depth or tree-based Wasserstein distance) (Urbani et al., 2024),
• Enforcing nested groupings in loss functions and evaluation.

2. Architectures and Losses for Hierarchical Supervision

Supervised approaches incorporate hierarchies directly into the model structure, output parameterization, or training losses.

a) Hierarchy-aware Losses

Architectural-agnostic methods extend cross-entropy by aggregating log probabilities over hierarchical ancestors of each label, weighted to balance fine- and coarse-grained accuracy. The resulting hierarchical loss is a proper scoring rule: its risk minimizer recovers the true posterior distribution over leaves (Urbani et al., 2024). Coarsening-focused losses reduce error rates on "gross" misclassifications (predictions far from the ground truth in the hierarchy).

b) Model Structures

Hierarchically structured classifiers use multi-headed heads, each dedicated to a specific hierarchy level or subtask (Garg et al., 2021, Aldabe et al., 2015). Some methods employ fine-to-coarse architectures: for each depth, a distinct classifier head receives as input both raw features and finer-level outputs, with gradient-control mechanisms to avoid information leakage (Garg et al., 2021). Binary SVM trees can hierarchically partition the label space using confusion-based clustering, decomposing multi-class classification as a decision path through a tree (Aldabe et al., 2015).

c) Graph-based Label Modeling

Graph convolutional networks (GCNs) or attention-based propagation enable explicit encoding of hierarchical label dependencies, such as in vision transformers with class tokens per hierarchy node (Stoimchev et al., 12 Mar 2026), or label embeddings processed through a graph-attention network (GAT) for taxonomy-informed projection (Yu et al., 2023).

3. Semi-Supervised and Weakly-Supervised Hierarchical Learning

Hierarchical supervision plays an essential role when high-quality annotation is expensive or unattainable.

a) Leveraging Weak and Coarse Labels

Methods such as HierMatch (Garg et al., 2021) exploit coarse-grained labels as weak supervision, allocating classifier capacity to all hierarchy levels. For instance, a sample labeled only at a superclass node is still fitted via cross-entropy at that node. During learning, classifiers at every depth are trained with a mixture of full (fine) supervision and weak (coarse) supervision; unlabeled data are incorporated via standard consistency or pseudo-labeling objectives, instantiated at each hierarchical granularity (Garg et al., 2021).

b) Multi-path and Multi-level Semi-supervision

Frameworks for semi-supervised hierarchical multi-label classification, such as HELM (Stoimchev et al., 12 Mar 2026), construct explicit models for both multi-path (multi-branch) label assignments and partial supervision from unlabeled data. These systems employ hierarchy-specific tokens, GCN-based propagation, and self-supervised objectives (e.g., BYOL branches) to jointly optimize supervised and hierarchy-aware losses using both labeled and unlabeled images.

Pseudo-labeling strategies for hierarchical open-set classification use subtree pseudo-labels: an unlabeled sample is assigned to a node if its total probability mass in that subtree exceeds a threshold, providing robust weak supervision without over-committing to leaves (critical for OOD data) (Wallin et al., 23 Jan 2026). Mechanisms such as age-gating prevent the proliferation of spurious deep pseudo-labels by filtering late-assigned, low-confidence labels.

c) Label Discovery under Hierarchical Supervision

When only coarse supervision is available, hierarchical contrastive or self-contrastive methods can discover fine-grained categories by transferring structure from shallow (coarse-supervised) to deep (fine-discovering) layers (An et al., 2022). Multistage objectives combine supervised cross-entropy at the coarse layer with weighted contrastive losses that differentiate within-coarse-class variation to separate fine-grained clusters post-hoc (e.g., via K-means).

4. Hierarchy-Aware Contrastive and Representation Learning

Contrastive learning, when naively implemented, assumes a flat label space, but several frameworks extend it to hierarchies (Kim et al., 2022, Zhang et al., 2022, Lian et al., 2024, Yu et al., 2023).

a) Hierarchical Multi-label Contrastive Losses

Losses are structured to aggregate positives and negatives along the hierarchy, possibly weighting finer-level positive pairs more heavily and enforcing a monotonicity constraint so that closeness in the hierarchy translates to proximity in embedding space (Zhang et al., 2022). These constraints produce learned representations in which fine-grained siblings cluster tightly under their coarse ancestor, and different families are well-separated, as measured by intra- and inter-cluster distances (Lian et al., 2024).

b) Proxy-based and Hyperbolic Regularization

Hierarchical proxies, learned as points in hyperbolic space (e.g., the Poincaré ball), can serve as virtual ancestors or centroids, enabling models to discover latent semantic hierarchies (Kim et al., 2022). Triplet-based losses pull together samples and their lowest-common-ancestor proxies while pushing apart unrelated groups, yielding emergent tree-like clustering without explicit tree supervision.

c) Supervised Contrastive Learning for Multilabel Hierarchies

Instance-wise and label-wise joint contrastive objectives exploit the full taxonomic structure: instance-wise terms organize the batch so that samples sharing labels at a chosen depth are pulled together, with deep-level agreement prioritized exponentially; label-wise terms cluster batchwise label embeddings, weighted by hierarchy-sensitive Hamming distances (Yu et al., 2023). These structures yield improved macro-F1 and more stability under multi-path settings.

5. Practical Aspects: Annotation, Continual Learning, and Taxonomy Enrichment

a) Hierarchical Annotation Interfaces

Crowdsourcing interfaces that present annotators with tasks organized by the label hierarchy deliver higher annotation quality and efficiency compared to flat or random grouping. Specifically, semantic grouping of labels increases macro-F1 scores, filtering out negatives at higher levels boosts precision, and full-hierarchy UIs aid on high-difficulty instances (Stureborg et al., 2023).

b) Continual Learning and Label Expansion

Hierarchical label expansion arises in continual learning, where initially coarse classes are incrementally refined into finer classes as the data stream progresses (Lee et al., 2023). Hierarchy-aware pseudo-labeling and memory management (rehearsal-based PL-FMS) selectively store and transfer samples to balance stability and plasticity, yielding significant improvements over conventional, flat continual learning benchmarks.

c) Minimal and Weak Supervision via Taxonomy Enrichment

In low-supervision scenarios, methods such as TELEClass combine LLM-driven generation and annotation with taxonomy enrichment (mining corpus-driven class-indicative phrases) to bootstrap hierarchical classifiers from only class names (Zhang et al., 2024). Document–class scoring is refined using semantic similarity in embedding space, and classifiers enforce multi-label and hierarchical constraints. This setup achieves competitive performance against zero-shot LLM prompting at a fraction of the inference cost.

6. Empirical Impact, Evaluation Metrics, and Open Issues

Hierarchical labeling and supervision consistently decrease semantically gross errors and improve prediction quality at both coarse and fine levels, often matching or exceeding flat baselines in pure accuracy and yielding further gains in hierarchical distance or error measures (such as Wasserstein cost on the label tree) (Urbani et al., 2024). Evaluations typically report standard metrics (e.g., micro/macro F1, accuracy, MAP@R), along with hierarchy-aware measures (LCA depth, mean hierarchical distance, hierarchical precision/recall).

Key considerations and limitations include:

• Dependence on the meaningfulness and depth of the hierarchy,
• Robustness in the presence of partial or noisy labels,
• Hyperparameter sensitivity in semi-supervised regimes (e.g., thresholds for pseudo-labeling, loss weighting),
• Calibration and interpretability of hierarchy-induced uncertainties.

Future directions include dynamic or learned weighting of hierarchy levels, better calibration techniques, adaptation to partial labels, and integration of taxonomy learning with model training. Hierarchically structured supervision remains a central avenue for scalable, efficient, and robust learning in structured prediction tasks across domains.