On the Transferability Between Extreme Multi-Label and Hierarchical Text Classification

This repository accompanies our study of whether established extreme multi-label classification (XML) and hierarchical text classification (HTC) methods transfer effectively to datasets from the other domain.

At a glance

• Read the paper
• Read the final results
• Convert datasets between HTC and XML formats
• Set up an individual model integration
• See authors and acknowledgements
• Cite the paper

The complete values, including deviations over runs, are presented in RESULTS.md.

Model integrations

All evaluated models were introduced by their original authors. This repository contains study-specific dataset conversion, adaptations, evaluation and the paper results.

Using the repository

There is no single environment for all five methods. The integrations retain different research-code dependency stacks and should be set up separately:

• CascadeXML setup and expected data
• XR-Transformer environment guide
• HBGL setup, preprocessing and launchers
• HGCLR setup and usage
• RADAr++ setup and usage
• HTC/XML dataset conversion

The repository does not provide a unified training API. It preserves the study integrations and the commands needed to understand and reuse the adaptations.

Data

The datasets are not redistributed as part of this repository. Users must obtain the original WOS, NYT, RCV1-V2, Wiki10-31K and AmazonCat-13K data under the terms of their respective providers.

The converters under dataset_transfer/ translate between the line-oriented XML representation and the JSON Lines HTC representation used by the model integrations. Generated datasets, checkpoints, caches and scheduler logs are excluded from Git.

Citation

If you use this repository, please cite:

Florian Hauss, Tom Speier, Nerijus Bertalis, Paul Granse, Ferhat Gül, Leon Menkel, David Schüler, Lukas Galke Poech, and Ansgar Scherp. On the Transferability Between Extreme Multi-Label and Hierarchical Text Classification. doi:10.1145/3820755.3832808.

Please also cite the original publication for each evaluated method. Upstream repositories, licences and attribution are listed in THIRD_PARTY_NOTICES.md.