florisdh/hcc

Hierarchical Conformal Classification Benchmarks

Three benchmarks for conformal prediction over hierarchical label spaces, released with the paper Hierarchical Conformal Classification (NeurIPS 2026). Each benchmark pairs the predicted class probabilities of a trained classifier with its leaf nodes, and provides a class taxonomy. The leaf-node prediction scores and taxonomy can be jointly used to create prediction sets of hierarchy nodes (rather than flat leaf labels) and evaluated without retraining any model.

Config	Modality	Source data	Classifier	Accuracy	Classes (leaves)	Taxonomy nodes	Depth	Hierarchy
dbp	text	DBpedia14 (Wikipedia abstracts)	BERT	.99	14	25	6	tree
img	image	ImageNet-1K validation set	ResNet-152	.79	1,000	1,860	17	DAG (75 nodes with multiple parents)
gtz	audio	GTZAN genre collection	XGBoost	.63	10	15	3	tree

Splits

Each config has a calibration and a test split: a stratified 20/80 split with random_state=42, identical to the split used in the paper.

Config	calibration	test
dbp	14,000	56,000
img	10,000	40,000
gtz	199	800

Data fields

• id (int): row identifier in the original prediction files.
• input (string): the input, or a pointer to it. dbp: the Wikipedia abstract. img: the ImageNet validation image file name. gtz: the index of the GTZAN track.
• label (ClassLabel): the true leaf class; class names are the taxonomy's synset identifiers.
• scores (list of float32): the classifier's predicted probabilities over all classes, in the order of the label class names.

Taxonomies

The taxonomies/ folder contains, per config:

• {config}_lexicon.xml: the taxonomy as a WordNet-LMF lexicon, loadable with the wn package;
• {config}_edges.csv: the same taxonomy as a parent,child edge list;
• {config}_id_to_label.json: the mapping from class index to taxonomy node.

The ImageNet taxonomy is the subset of WordNet spanned by the 1,000 ImageNet classes and their ancestors. The DBpedia14 taxonomy is derived from the DBpedia ontology and the GTZAN taxonomy follows Sturm (2012).

Usage

First install the wordnet package

pip install wn

Now load the datasets' taxonomies

import wn
from huggingface_hub import hf_hub_download

path = hf_hub_download("your-org/hierarchical-conformal-classification",
                       "taxonomies/img_lexicon.xml", repo_type="dataset")
wn.add(path)
taxonomy = wn.Wordnet("omw-subset-en")

Now load the dataset

from datasets import load_dataset

ds = load_dataset("your-org/hierarchical-conformal-classification", "img")
calibration, test = ds["calibration"], ds["test"]
scores, label = test[0]["scores"], test[0]["label"]

Code to reproduce the paper's experiments is available at https://github.com/your-org/hierarchical-conformal-classification.

Attribution

Please cite

@inproceedings{denhengst2026hierarchical,
  title     = {Hierarchical Conformal Classification},
  author    = {den Hengst, Floris and Blin, In{\`e}s and Mohammadi, Majid and
               Shah, Syed Ihtesham and Younesian, Taranah},
  booktitle = {Advances in Neural Information Processing Systems},
  publisher = {Curran Associates, Inc.},
  volume = {39, Main Conference},
  year = {2026}
}