Taxonomy-Adaptive Classification

• Taxonomy-adaptive classification is a method that actively uses hierarchical taxonomies to shape label definition, data organization, and model design across various domains.
• It employs regimes like hierarchical target prediction and cross-taxonomy alignment to enhance performance, calibration, and interpretability in complex classification tasks.
• Concrete examples include applications in textiles, patent mapping, and metagenomic analysis, where adaptive taxonomies lead to improved accuracy and computational efficiency.

Searching arXiv for recent and foundational papers on taxonomy-adaptive classification across domains. Taxonomy-adaptive classification denotes a family of classification procedures in which a taxonomy is not treated as a passive label inventory but as an active source of inductive bias for label definition, data organization, model design, inference, and evaluation. In the literature, this role appears in several technically distinct forms: selecting taxonomy levels as prediction targets in robotics and metagenomics, propagating evidence over hierarchical label graphs in patent and product classification, regularizing flat classifiers with taxonomic constraints, reconciling mismatched label spaces across datasets and domains, and expanding taxonomies as corpora evolve (Longhini et al., 2021, Fuhl et al., 2023, Zou et al., 2023, Chen et al., 12 Jan 2025, Gong et al., 2021, Kargupta et al., 12 Jun 2025). The common principle is that classification performance, calibration, and interpretability improve when the learned decision structure is aligned with the hierarchical organization of the phenomenon being classified rather than with a flat or taxonomically incomplete label set.

1. Problem regimes and conceptual scope

Taxonomy-adaptive classification is not a single algorithmic template. The surveyed literature uses the term for several regimes that share the same design principle: the taxonomy governs what distinctions are learned and how predictions are made.

Regime	Taxonomy role	Representative work
Hierarchical target prediction	Parent–child structure constrains label prediction	(Zou et al., 2023, Chen et al., 12 Jan 2025)
Rank-adaptive classification	Different taxonomy levels use different features/models	(Fuhl et al., 2023, Su et al., 2021)
Cross-taxonomy alignment	Different datasets or domains are reconciled through a common taxonomy	(Fillies et al., 7 Mar 2025, Gong et al., 2021, Saha et al., 2015)
Evolving taxonomy	New classes or branches appear over time	(Lake, 2022, Kargupta et al., 12 Jun 2025)
Taxonomy as prior in flat classification	Hierarchy regularizes a leaf-level classifier	(Pourvali et al., 2023)
Taxonomy over domains	Domain relations, not output labels, are hierarchical	(Liu et al., 2023)

In hierarchical output settings, the taxonomy defines admissible label paths. Patent classification models the International Patent Classification as a graph with intra-level sibling edges and inter-level parent–child edges, then predicts at $L3$ using taxonomy-aware code embeddings and assignee histories (Zou et al., 2023). Multimodal e-commerce classification encodes transitions between adjacent levels with binary matrices $M^{[\ell_i,\ell_i+1]}$, so deeper predictions are multiplicatively conditioned on higher-level probabilities (Chen et al., 12 Jan 2025).

In rank-adaptive settings, the taxonomy determines which level should be predicted and which signals should be used. In metagenomics, separate classifiers are trained for superkingdom, phylum, and genus, with $k$-mer dimensionality and model family tuned per rank; a 4-mer subspace $k$-NN gives the best genus MAP of $0.8643$ on the final dataset, while other ranks prefer different $k$ and model choices (Fuhl et al., 2023). In semi-supervised biodiversity recognition, a single species-level head is trained while coarse taxonomic labels such as Kingdom, Phylum, Class, or Order are incorporated through marginalization from species probabilities to coarser levels (Su et al., 2021).

A different regime arises when taxonomies do not match across datasets. Cross-taxonomy hate-speech classification constructs a universal five-layer taxonomy and trains a single multi-label classifier over that shared label space, improving out-of-domain ETS F1 from $0.73$ to $0.84$ after integration of a second dataset (Fillies et al., 7 Mar 2025). TACS addresses semantic segmentation when source and target taxonomies differ, including open taxonomy, coarse-to-fine taxonomy, and implicitly-overlapping taxonomy (Gong et al., 2021). Earlier work on taxonomy-grounded aggregation addresses a related but weaker setting: multiple pre-trained classifiers with different label sets are grounded into a common class taxonomy and aggregated through heuristic score propagation or a graphical model (Saha et al., 2015).

2. Taxonomy as target space, routing structure, and decoding constraint

A central mechanism is to let the taxonomy determine the prediction target itself. This can be done without explicit hierarchy-aware loss functions, simply by choosing taxonomy levels as labels and collecting data that makes those levels mechanically or semantically salient. Textile manipulation provides a clear case. The taxonomy has two main branches—fiber material and production method—and the experiments show that production method dominates force–torque response under pulling and twisting. Construction-based labels yield markedly higher accuracy than material-only labels, with the CNN reaching $100\%$ on $D^{pull}$ for construction and $M^{[\ell_i,\ell_i+1]}$0 for material; the active-arm SVM reaches $M^{[\ell_i,\ell_i+1]}$1 construction accuracy with all pull signals and with $M^{[\ell_i,\ell_i+1]}$2 alone, while material accuracy is lower (Longhini et al., 2021). The key adaptation is not a taxonomy-aware loss but the choice of classification target and of actions that expose the target’s mechanics.

In metagenomics, taxonomy-adaptive classification is explicitly rank-specific. Features are $M^{[\ell_i,\ell_i+1]}$3-normalized $M^{[\ell_i,\ell_i+1]}$4-mer histograms,

$M^{[\ell_i,\ell_i+1]}$5

and the classifier is chosen per taxonomic rank rather than globally. On the final dataset, subspace $M^{[\ell_i,\ell_i+1]}$6-NN with 4-mer features is best for genus; on the non-similar dataset, bagged trees with 3-mers are best for phylum, and subspace $M^{[\ell_i,\ell_i+1]}$7-NN with 3-mers is best for superkingdom (Fuhl et al., 2023). This is taxonomy-adaptive because separability varies by rank.

A stronger form of taxonomic routing appears in transitional hierarchical classifiers. In multimodal product categorization, the taxonomy is encoded by fixed transition matrices. Given logits $M^{[\ell_i,\ell_i+1]}$8 and level-wise probabilities $M^{[\ell_i,\ell_i+1]}$9, the next-level structural attention is

$k$0

and the next prediction is

$k$1

This yields deeper predictions that are explicitly conditioned on higher-level mass and improves Consistency across every evaluated backbone; for example, on Llama (7B), Consistency rises from $k$2 to $k$3, Acc@$k$4 rises from $k$5 to $k$6, and Exact Match rises from $k$7 to $k$8 (Chen et al., 12 Jan 2025).

Patent classification uses a graph-based variant of the same idea. IPC codes are nodes in a graph $k$9 with sibling and parent–child edges, and code embeddings are updated by adaptive attention over both neighborhoods. The final prediction combines taxonomy-aware text attention and assignee behavior: $k$0 This joint taxonomy-and-history formulation yields gains up to $k$1 precision, $k$2 recall, and $k$3 NDCG over the best baseline on USPTO-200K, and up to $k$4, $k$5, and $k$6 on CNPTD-200K (Zou et al., 2023).

3. Taxonomy as inductive bias in losses, priors, and adaptation objectives

Another major line of work keeps the prediction task flat or partially flat but injects the taxonomy into the optimization criterion. In fine-grained semi-supervised learning, the model predicts species probabilities $k$7, and coarser probabilities are obtained by marginalization,

$k$8

The hierarchical supervised loss combines species supervision for labeled images and coarse-taxonomy supervision for coarsely labeled images,

$k$9

On Semi-iNat, incorporating Phylum labels improves species-level accuracy by $0.8643$0 in a transfer learning setting, and adding the hierarchical label structure with FixMatch improves performance by a further $0.8643$1 (Su et al., 2021).

TaxoKnow goes further by adding the taxonomy as an explicit regularizer to a flat multi-class classifier. One variant compiles hierarchical constraints into an SDD and penalizes the negative log weighted model count of taxonomy-consistent assignments; another uses a batch-based GCN teacher and aligns classifier outputs $0.8643$2 to taxonomy-aware outputs $0.8643$3 through

$0.8643$4

Across four datasets, the regularizers consistently improve Macro-F1. On Shopping Mall, supervised Macro-F1 rises from $0.8643$5 to $0.8643$6 with the GCN-based regularizer; on Reuters with $0.8643$7 labeled data, Macro-F1 rises from $0.8643$8 to $0.8643$9 (Pourvali et al., 2023).

Taxonomy-structured domain adaptation shifts the taxonomy from the output space to the domain space. Domains are organized by a tree-induced distance matrix $k$0, and the objective is

$k$1

where the taxonomist $k$2 regresses pairwise domain distances from representations. The analysis shows that with a non-informative flat taxonomy, the method reduces to classic adversarial domain adaptation, while informative taxonomies prevent full uniform alignment (Liu et al., 2023). This is a different sense of taxonomy-adaptation, but it preserves the same principle: hierarchical structure governs which invariances should and should not be learned.

Bayesian nonparametric taxonomy adaptation provides a probabilistic variant. BayesANT places species sampling priors at each taxonomic rank so that new taxa can be discovered. The leaf posterior has the generic form

$k$3

with “new” allocations available at every rank. On FinBOL, species accuracy reaches $k$4 in one scenario and $k$5 in a harder scenario, while preserving explicit probabilities for novelty (Zito et al., 2022).

4. Cross-taxonomy integration, evolving taxonomies, and open-world adaptation

Taxonomy-adaptive classification is especially important when the taxonomy itself changes. Zero-shot job classification treats taxonomy evolution as multi-label document classification over an evolving label set. Instead of per-class parameters $k$6, the model scores a document $k$7 against the text of any class $k$8 by a semantic compatibility function $k$9. Under a controlled training budget, a cross-encoder zero-shot model achieves LRAP $0.73$0 and macro-AP $0.73$1, compared with $0.73$2 and $0.73$3 for a traditional multi-label classifier trained on all classes, corresponding to $0.73$4 LRAP and $0.73$5 macro-AP (Lake, 2022). This directly addresses taxonomy expansion scenarios such as Refine and Extend.

TaxoAdapt addresses a different form of taxonomy evolution: the taxonomy must be grown to reflect an evolving scientific corpus. For each dimension $0.73$6, node density is $0.73$7, and non-leaf unmapped density is

$0.73$8

Depth expansion is triggered for leaves with $0.73$9; width expansion is triggered for non-leaves with $0.84$0. With $0.84$1 and $0.84$2, TaxoAdapt produces multidimensional taxonomies that are $0.84$3 more granularity-preserving and $0.84$4 more coherent than the strongest baselines, while also increasing dimension specificity, relevance, and coverage (Kargupta et al., 12 Jun 2025). This suggests that taxonomy-adaptive classification can be inseparable from taxonomy construction when the domain is rapidly changing.

Open-world and mismatched-taxonomy settings require explicit alignment operators. TACS introduces stochastic label mapping for one-to-many source-to-target relations, pseudo-label based relabeling, bilateral mixed sampling, and uncertainty-rectified contrastive learning. The framework is evaluated on open taxonomy, coarse-to-fine taxonomy, and implicitly-overlapping taxonomy scenarios and substantially improves mIoU over prior work; for example, in the open-taxonomy SYNTHIA $0.84$5 Cityscapes setting, the method reaches $0.84$6 on the novel-6 classes and $0.84$7 overall (Gong et al., 2021). Cross-taxonomy hate-speech integration follows a related logic at the document level by building a universal five-layer taxonomy and mapping source labels deterministically into that space, reducing dependence on multiple specialized classifiers (Fillies et al., 7 Mar 2025).

Taxonomy grounding also supports model aggregation without retraining. A heuristic roll-up method and a mixed discrete-Gaussian graphical model aggregate predictions from classifiers with different label sets by mapping them into a common taxonomy and enforcing hierarchical consistency. On image classification, the graphical model reaches $0.84$8, $0.84$9, and $100\%$0, and with entry-level termination reaches $100\%$1, $100\%$2, and $100\%$3 (Saha et al., 2015).

5. Empirical regularities, interpretability, and systems trade-offs

Several empirical regularities recur across the literature. First, taxonomy-aligned targets often outperform semantically flatter targets. Textile classification by production method is easier than classification by fiber content because construction dominates the measured force–torque response (Longhini et al., 2021). Metagenomic classification is easier at superkingdom than at genus, and this motivates rank-specific feature and model choices (Fuhl et al., 2023). In hierarchical product classification, deeper-level accuracy improves when predictions are conditioned on parent distributions rather than produced by independent heads (Chen et al., 12 Jan 2025).

Second, hierarchy-aware methods often improve minority or sparse-label behavior more than head-class performance. TaxoKnow’s most consistent gains are in Macro-F1 rather than only Accuracy, and semi-supervised hierarchy-aware learning in Semi-iNat yields larger relative gains when finer taxonomic labels such as Class or Order are provided or when models are trained from scratch (Pourvali et al., 2023, Su et al., 2021). This suggests that taxonomies primarily regularize underdetermined regions of the label space.

Third, taxonomy-adaptation is frequently tied to interpretability. In textiles, t-SNE, channel-wise statistics, and Grad-CAM reveal that woven cotton drives strong $100\%$4, $100\%$5, and $100\%$6 attention, whereas knitted cotton draws broader force-channel and $100\%$7 attention, matching construction-dependent mechanics (Longhini et al., 2021). Patent IPC embeddings visualized by t-SNE form coherent clusters corresponding to technology fields and subclusters aligned with co-occurrence structure (Zou et al., 2023).

Fourth, taxonomy-aware methods can be computationally frugal when the hierarchy is used to reduce search. In metagenomics, feature-space balancing reduces the training set to approximately $100\%$8 on the non-similar dataset and approximately $100\%$9 on the final dataset while often improving MAP; bagged trees run at $D^{pull}$0 s per sequence with $D^{pull}$1 GB, and NN-512 at $D^{pull}$2 s with $D^{pull}$3 GB (Fuhl et al., 2023). In zero-shot job classification, a bi-encoder filter followed by a cross-encoder re-ranker cuts forward-pass overhead by approximately $D^{pull}$4 with only a small relative LRAP decrease (Lake, 2022).

Benchmark construction is beginning to expose failure modes systematically. MMM-Bench provides a five-level taxonomy over 12 commercial domains and shows large performance drops from $D^{pull}$5 to $D^{pull}$6–$D^{pull}$7, long-tail degradation from $D^{pull}$8 Macro-F1 on head $D^{pull}$9 classes to $M^{[\ell_i,\ell_i+1]}$00 on tail classes, and modality ablations in which text-only SFT can outperform multimodal inputs, indicating cross-modal utilization failure rather than a generic benefit from more modalities (Ma et al., 11 May 2026).

6. Limitations and open directions

The literature is technically diverse but converges on several unresolved problems. Scalability is a recurring concern. Textile experiments cover three fiber classes, two constructions, and 40 samples, and explicitly leave generalization to more taxa, finishing processes, blends, and non-woven constructions open (Longhini et al., 2021). The metagenomic study fixes read length at 1500 nt and does not separately model open-set detection beyond an “unknown” class (Fuhl et al., 2023). Multimodal transitional classifiers note that transition-matrix construction becomes increasingly complex for large and deep hierarchies and that sequential top-down inference reduces parallelizability (Chen et al., 12 Jan 2025).

Another limitation is taxonomy quality itself. TaxoKnow assumes a clean tree structure; extending to DAG taxonomies requires revised constraints and compilation strategies (Pourvali et al., 2023). TSDA is only advantageous when the taxonomy is informative; with a flat taxonomy it provably reduces to the standard adversarial solution (Liu et al., 2023). TaxoAdapt depends on LLM judgments for dimension classification, pseudo-labeling, clustering, and metric computation, and sparse areas of the corpus may be under-expanded when $M^{[\ell_i,\ell_i+1]}$01 is too high (Kargupta et al., 12 Jun 2025).

Open-world behavior remains uneven. BayesANT provides explicit probabilities for unseen taxa (Zito et al., 2022), and TACS addresses open taxonomy by using few-shot target labels plus mapping and relabeling (Gong et al., 2021). By contrast, several other systems assume closed-world taxonomies or rely on “unknown” as an ordinary class rather than as a separately modeled open-set process (Fuhl et al., 2023). This suggests that taxonomy-adaptive classification and open-set recognition are still only partially integrated.

A further limitation is that hierarchical consistency does not automatically imply semantically correct fine-grained discrimination. MMM-Bench shows that even strong models retain large gaps at $M^{[\ell_i,\ell_i+1]}$02–$M^{[\ell_i,\ell_i+1]}$03, and the benchmark argues for hierarchy-aware objectives, constrained decoding, label semantics, and domain-adaptive fusion as still-missing ingredients (Ma et al., 11 May 2026). In biodiversity multimodal learning, hierarchy-guided representation learning improves taxonomic classification accuracy by over 14 percent under corrupted DNA conditions, but the reported setting remains closed-set and does not yet solve the seen/unseen split inherited from earlier benchmarks (Ahmed et al., 26 Mar 2026).

The cumulative evidence therefore supports a precise but limited conclusion. Taxonomy-adaptive classification works when the taxonomy is treated as a model-bearing object: a source of admissible paths, a guide to which distinctions matter, a scaffold for regularization, a basis for cross-dataset alignment, or a trigger for adaptive expansion. It does not eliminate the need for high-quality mappings, adequate supervision at critical levels, or careful handling of open-world drift. What the literature consistently shows is narrower and more technical: when a classification problem is intrinsically hierarchical, models that ignore that hierarchy learn the wrong geometry, waste supervision, and often fail exactly where taxonomies are most informative.