pyena 0.2.0

Python implementation of Epistemic Network Analysis, verified against rENA

pyena

pyena is a Python implementation of Epistemic Network Analysis (ENA), numerically verified against the reference R package rENA 0.3.1.

Features

• Verified adjacency vectors — reverse-engineered from rENA's C++ source (vector_to_ut + ref_window_df), using the cross-speaker moving stanza window. Verified exactly against rENA on nine datasets (RS.data + eight synthetic).
• SVD projection + means rotation — both stages match rENA outputs at machine epsilon on RS.data (RMSE < 1e-16).
• Group comparison — Welch's t, Mann-Whitney U, and permutation test in one call, with automatic small-sample diagnostics.
• Reproducibility metrics — four indicators (centroid dispersion, pairwise distance, 95% confidence ellipse area, convex hull area) for quantifying the spread of repeated analyses under nominally identical conditions.
• scikit-learn-style API — single ENA estimator that orchestrates the full pipeline.

Installation

pip install pyena

Requires Python 3.10+.

Development install

git clone https://github.com/zhegllyang/pyena.git
cd pyena
pip install -e ".[dev]"

Quickstart

import pandas as pd
from pyena import ENA

df = pd.read_csv("your_long_format_utterances.csv")

ena = ENA(
    codes=["Data", "Theory", "Question", "Example", "Critique"],
    unit_col="unit",
    conversation_col="conversation",
    window_size=4,
    rotation="means",
    mr_groups=("A", "B"),
)
ena.fit(df)

ena.plot()
result = ena.compare(axis="x")
metrics = ena.reproducibility()

Validation against rENA 0.3.1

Stage	Test data	Match	Max abs diff
Adjacency vectors	nine datasets (RS.data + 8 synthetic)	all cells	0
SVD coordinates	RS.data (48 units × 2 axes)	96 / 96	< 1e-15
Means rotation	RS.data (48 units × 2 axes)	96 / 96	< 1e-15
Node placement	RS.data (6 codes × 2 axes)	12 / 12	< 1e-13

See tests/test_adjacency.py, tests/test_projection.py, and tests/test_crossspeaker.py for the regression tests, and the R scripts in validation_rsdata/ for regenerating the rENA reference outputs.

Testing

pytest

108 tests cover reference regression, mathematical properties of each algorithm, and external cross-checks against scipy (Welch, Mann-Whitney) and the rENA reference outputs.

Citation

If you use pyena in your research, please cite:

@software{song_pyena_2026,
  author    = {Song, JongHwi},
  title     = {{pyena: Python implementation of Epistemic Network Analysis, verified against rENA}},
  year      = {2026},
  publisher = {Zenodo},
  version   = {0.2.0},
  doi       = {10.5281/zenodo.20339527},
  url       = {https://github.com/zhegllyang/pyena}
}

A SoftwareX paper describing pyena is in preparation.

License

BSD 3-Clause. See LICENSE.

Acknowledgements

pyena reimplements the algorithms of rENA 0.3.1 by Marquart, Swiecki, Collier, Eagan, Woodward, and Shaffer. The reimplementation was done independently from the publicly available C++ and R source code.