Release Notes & Changelog

Version history and changelog for poisson-topicmodels.

Current Version

Latest: 0.2.0 (April 2026)

For full changelog, see below.

Version 0.2.0 - Inspection & Diagnostics

April 1, 2026

New Methods (all models):

  • ✨ summary(n_top_words=5) — Formatted text summary of any fitted model

  • ✨ compute_topic_coherence(metric='c_npmi', top_n=10) — Per-topic NPMI/UMass coherence

  • ✨ compute_topic_diversity(top_n=25) — Unique-word fraction across topics (0–1)

  • ✨ plot_topic_prevalence() — Horizontal bar chart of mean topic prevalence

  • ✨ plot_topic_correlation() — Cosine-similarity heatmap between topics

  • ✨ plot_document_topic_heatmap() — Document × topic heatmap

  • ✨ Academic-style plotting (_setup_academic_style) applied to all built-in figures

New Methods (SPF):

  • ✨ plot_seed_effectiveness() — Grouped bar chart of seed vs. non-seed word weights

New Methods (CPF & CSPF):

  • ✨ return_covariate_effects_ci(ci=0.95) — Covariate effects with Bayesian credible intervals

  • ✨ plot_cov_effects(ci=0.95) — Forest plot of covariate effects

New Methods (TBIP):

  • ✨ return_ideal_points() — DataFrame of author positions with uncertainty

  • ✨ return_ideological_words(topic, n) — Top ideological words per topic

  • ✨ plot_ideal_points(show_ci=True) — Publication-ready 1-D scatter with CIs

  • ✨ return_topics() / return_beta() overrides using LogNormal posterior

New Methods (ETM):

  • ✨ return_topics() — Neural encoder inference (was NotImplementedError)

  • ✨ return_beta() — Embedding-based topic–word computation (was NotImplementedError)

New Model:

  • ✨ STBS (Structured Text-Based Scaling) — topic-specific ideal points with author-level covariate effects

Bug Fixes:

  • 🐛 TBIP: Fixed sigma_y → sigma_x typo in variational guide

  • 🐛 CPF: Fixed self.covariates not set in __init__; fixed ndim validation order

  • 🐛 PF: Removed duplicate return_top_words_per_topic override

Breaking Changes:

  • CSPF2 renamed to CSPF (old CSPF class removed)

  • from poisson_topicmodels import CSPF2 → from poisson_topicmodels import CSPF

Documentation:

  • All code examples updated to use correct method names (return_topics, return_beta, return_top_words_per_topic, train_step instead of fictional get_*/train names)

  • API reference rewritten with complete method documentation

  • Autodoc directives enabled for auto-generated class reference

  • New methods documented across fundamentals, tutorials, and examples

Version 0.1.0 - Initial Release

February 5, 2026

New Features:

  • ✨ Poisson Factorization (PF) - Unsupervised topic modeling

  • ✨ Seeded PF (SPF) - Guided topic discovery with keyword priors

  • ✨ Covariate PF (CPF) - Model topic variation by document metadata

  • ✨ Covariate Seeded PF (CSPF) - Combine seeds and covariates

  • ✨ Text-Based Ideal Points (TBIP) - Estimate author positions

  • ✨ Embedded Topic Models (ETM) - Integration with pre-trained embeddings

  • ✨ Stochastic Variational Inference (SVI) with mini-batch training

  • ✨ GPU acceleration via JAX

  • ✨ Comprehensive type hints (90%+ coverage)

  • ✨ >70% test coverage

  • ✨ Complete documentation and tutorials

Documentation:

  • Getting Started guide

  • Fundamentals covering all models

  • 4 detailed tutorials

  • How-to guides for common tasks

  • Complete API reference

  • Testing guide

  • Contributing guidelines

Quality:

  • Code follows Black, isort, mypy standards

  • Comprehensive error messages

  • Input validation

  • Reproducibility via seeding

  • GitHub Actions CI/CD

Breaking Changes:

  • N/A (first release)

Migration Guide:

  • N/A (first release)

Contributors:

  • Bernd Prostmaier (Lead)

  • Bettina Grün

  • Paul Hofmarcher

Known Issues:

  • GPU memory estimation could be improved

  • Some edge cases in covariate handling

  • Documentation for advanced features could be more detailed

Future Roadmap:

  • Dynamic topic models (time-varying topics)

  • Online learning from streaming data

  • Hierarchical topic models (HTM)

  • Better visualization toolkit

  • Performance optimizations

  • Additional embedding support

Coming Soon (0.3.0)

Planned features:

  • Dynamic Topic Models: Topics that evolve over time

  • Streaming Mode: Learn from new documents continuously

  • Better Visualization: Interactive topic visualization

  • Save/Load: Persist trained models to disk

  • Documentation Enhancements: More examples and tutorials

Not Planned for 0.2.0:

  • Breaking API changes (we’ll maintain compatibility)

  • Complete rewrite of inference (current approach is solid)

Deprecation Policy

Stability Guarantees:

  • Public API stable across minor versions

  • Breaking changes only in major versions

  • Deprecations announced one release ahead

How to stay updated:

  • Watch GitHub releases

  • Subscribe to changelog

  • Monitor upgrade guides

Installation by Version

Install specific version:

pip install poisson-topicmodels==0.1.0

List available versions:

pip index versions poisson-topicmodels

Upgrade to latest:

pip install --upgrade poisson-topicmodels

Compatibility Matrix

poisson-topicmodels

Python 3.11

Python 3.12

Python 3.13

0.1.0

✓

✓

✓

Dependency Versions

Core dependencies for 0.1.0:

  • jax==0.8.0

  • jaxlib==0.8.0

  • numpyro==0.19.0

  • numpy>=2.2.0,<3.0.0

  • scipy>=1.15.0,<2.0.0

  • pandas>=2.2.0,<3.0.0

  • scikit-learn>=1.6.0,<2.0.0

  • matplotlib>=3.10.0,<4.0.0

Optional dependencies:

  • sphinx>=6.0 (for documentation)

  • sphinx-rtd-theme>=1.2 (for docs theme)

  • myst-parser>=1.0 (for markdown docs)

  • pytest>=9.0.1 (for testing)

Citing poisson-topicmodels

If you use poisson-topicmodels in research, please cite:

@software{prostmaier2026poisson,
  title={poisson-topicmodels: Probabilistic Topic Modeling with Bayesian Inference},
  author={Prostmaier, Bernd and Grün, Bettina and Hofmarcher, Paul},
  year={2026},
  url={https://github.com/BPro2410/topicmodels_package}
}

Or in plain text:

Prostmaier, B., Grün, B., & Hofmarcher, P. (2026). poisson-topicmodels: Probabilistic Topic Modeling with Bayesian Inference. Retrieved from https://github.com/BPro2410/topicmodels_package

Community Feedback

Help improve poisson-topicmodels:

  • 🐛 Bug reports: GitHub Issues

  • 💡 Feature ideas: GitHub Discussions

  • 📚 Documentation feedback: GitHub Issues

  • 💬 General discussion: GitHub Discussions

  • 🤝 Contributions: See Contributing Guide

Report Issues

Found a bug? Please report it:

  1. Check existing issues

  2. Create detailed bug report

  3. Include: - Python version - OS - JAX version - Minimal reproducible example - Full error traceback

See Contributing Guide for detailed instructions.

Acknowledgments

Core Team:

  • Bernd Prostmaier

  • Bettina Grün

  • Paul Hofmarcher

Built on:

  • JAX - Automatic differentiation

  • NumPyro - Probabilistic programming

  • SciPy - Scientific computing

  • Sphinx - Documentation

Inspired by:

  • Gensim

  • scikit-learn

  • PyMC

  • NumPyro

License

poisson-topicmodels is licensed under the MIT License.

See LICENSE for details.

Next Steps

Questions? Open a discussion or create an issue on GitHub!

Questions About Releases

Q: When is the next release?

A: Quarterly minor releases, monthly patch releases. Check milestones on GitHub.

Q: How do I get a feature in the next release?

A: Open an issue or discussion to propose it, or submit a PR!

Q: Can I use development version?

A: Yes: pip install git+https://github.com/BPro2410/topicmodels_package.git

Q: What about backwards compatibility?

A: Public API stable within major versions. Breaking changes disclosed ahead of time.