ena-python 0.2.1

Python implementation of Epistemic Network Analysis, ported from the R package rENA.

ena-python

ena-python is a Python implementation of Epistemic Network Analysis, ported from the R package rENA.

Formerly pyENA. Renamed because an unrelated package already owns pyena on PyPI — including the pyena module name — so keeping it would have collided for anyone who installed both.

It is a standalone library: import ena_python needs only NumPy and pandas. No SciPy, no R, no Node, no server, and no network access. That keeps it usable in a notebook, in a script, or in browser Python — see the runnable Pyodide demo, which installs ena-python from PyPI and runs a full analysis in a browser tab.

Status: early release. Every numeric path — accumulation across all window types and models, the SVD / mean / generalized / regression / hENA rotations, node positions, variance/eigenvalues, Cohen's d, and the correlation functions — is checked against real compiled rENA 0.3.1; see the parity table, which also records the one deliberate, documented divergence. The API may still change.

Install

pip install ena-python

import ena_python

The distribution is ena-python, the module is ena_python, and the command is ena-python.

Optional extras:

pip install "ena-python[plot]"     # Plotly figures
pip install "ena-python[web]"      # FastAPI service
pip install "ena-python[parquet]"  # .parquet input

Or from source:

pip install git+https://github.com/HUDongpin/ena-python

Quick start

import pandas as pd
from ena_python import ena

rows = pd.DataFrame(
    {
        "UserName": ["u1", "u1", "u2", "u2"],
        "Condition": ["A", "A", "B", "B"],
        "GroupName": ["g1", "g1", "g2", "g2"],
        "Data": [1, 0, 1, 1],
        "Design": [0, 1, 1, 0],
        "Collaboration": [1, 1, 0, 1],
    }
)

set_ = ena(
    data=rows,
    codes=["Data", "Design", "Collaboration"],
    units=["Condition", "UserName"],
    conversation=["Condition", "GroupName"],
)

print(set_.points)      # unit positions in ENA space
print(set_.variance)    # variance explained, per dimension

Step by step

from ena_python import accumulate, make_set

data = accumulate(
    rows,
    units=["Condition", "UserName"],
    conversation=["Condition", "GroupName"],
    codes=["Data", "Design", "Collaboration"],
)
set_ = make_set(data, dimensions=2)

payload = set_.to_dict()  # JSON-friendly: NaN -> None, NumPy scalars -> Python scalars

to_dict() does not include your input dataset. ENA is often run over discourse transcripts that carry personal data, so results do not echo the source back by default. Pass to_dict(include_raw=True) if you want raw and row_connection_counts in the payload.

Input data

One row per coded line. Columns:

Column kind	Meaning
units	Who/what the network is about (e.g. Condition, UserName)
conversation	Which lines can connect to each other
codes	Binary 0/1 columns, one per code
metadata	Optional; carried through to results

Metadata must be constant within a unit. A metadata column that varies inside a unit has no unit-level value, so it is dropped and a UserWarning names it.

Interpreting the output

• set_.points — unit coordinates for the retained dimensions.
• set_.variance — variance explained for every dimension, normalized by the total across all of them (matching rENA). The first two entries sum to less than 1 whenever the data has rank > 2.
• set_.rotation.eigenvalues — prcomp sdev² (s²/(n−1)) for every dimension, as in rENA.
• set_.nodes — code node positions.

SVD signs are not canonical. As in any SVD, an axis may be mirrored relative to rENA or across platforms. Compare sign-aligned (see _sign_align in tests/test_r_oracle_parity.py).

A mean rotation's direction is arbitrary too. Passing rotation_params=[g1, g2] does not orient the axis from g1 toward g2 — swapping them yields an identical MR1, because the mean difference goes through a QR decomposition whose sign convention ignores the input's sign. rENA does the same. Read the group centroids to see which side is which rather than trusting the sign.

In the browser

ena-python runs under Pyodide with no server and no build step — the wheel is pure Python (a ~53 KB download) and its only dependencies (numpy, pandas) both ship as Pyodide packages.

const pyodide = await loadPyodide();
await pyodide.loadPackage(["micropip", "numpy", "pandas"]);
await pyodide.pyimport("micropip").install("ena-python");
pyodide.runPython(`from ena_python import ena; ...`);

examples/pyodide/ is a complete working page that does this and plots the result. On a 2026 laptop the analysis takes 0.02–0.07 s; the ~10–16 s cold start is almost entirely the browser downloading numpy and pandas as WebAssembly, which it then caches — about 7 MB, down from 20 MB before SciPy was dropped.

Plotting and serving

from ena_python.plotting import add_network, add_nodes, add_points, ena_plot

fig = ena_plot(set_)
add_network(fig)
add_points(fig)
add_nodes(fig)
fig.to_json()

uvicorn ena_python.web.api:app --reload   # /accumulate, /model, /ena, /plot

The web app is for localhost/trusted use: no authentication, and a row cap (ENA_PYTHON_WEB_MAX_ROWS, default 100000) is its only DoS guard. Put your own auth and limits in front of it before exposing it.

CLI

ena-python inspect examples/cli_sample.csv

ena-python ena examples/cli_sample.csv \
  --units unit --conversation conv --codes A B C \
  --metadata group score --window-size-back 2 \
  --output ena.json

Reads CSV, TSV, and Parquet (Parquet needs ena-python[parquet]). --include-raw echoes the input into the JSON. See docs/cli_usage.md.

Parity with rENA

Golden fixtures are generated from real, compiled rENA 0.3.1 and stamped with provenance; the test suite refuses a fixture built from anything less authoritative. Be aware of what is and isn't covered:

Area	Checked against rENA?
EndPoint accumulation, moving-stanza window	Yes — incl. infinite windows vs the compiled C++ kernel
Conversation window	Yes — accumulation and full model
AccumulatedTrajectory / SeparateTrajectory	Yes — accumulation and full model
Sphere/skip normalization, centering	Yes
SVD rotation, points, node positions, centroids	Yes (sign-aligned)
variance, eigenvalues	Yes — incl. a rank-3 case where a truncated denominator is wrong
Generalized rotation (gmr)	Yes — numeric and categorical targets, and x+y
hENA regression / regression_2	Yes for the x axis; y axis diverges by design (see below)
hENA rotation_h	Yes — incl. control variables
Mean rotation	Yes — full basis incl. the SVD completion (its direction is QR-conventional, as in rENA — see above)
Cohen's d	Yes — 6 cases, incl. mirrored inputs
ena_correlations, ena_correlation	Yes — pearson/spearman and the Fisher-CI kernel

See docs/testing_strategy.md and reference/README.md.

Issues found in rENA 0.3.1

Porting turned up six places where rENA 0.3.1 behaves differently from what its own code, comments, or docs intend. They are written up with reproducible snippets in docs/rena-upstream-issues.md, and tracked here under the upstream-rena label.

#	Issue	Severity	ena-python
1	ena.rotate.by.hena.regression does not deflate the y axis, returning axes ~0.8–0.97 collinear	High — affects coordinates	Deflates; axes orthogonal
2	Regression axes named after the first edge (A & B_reg), not the predictor; duplicate names with x+y	Low	Names after the predictor
3	The documented "lm(formula=V ~ ...)" param form errors	Low	Takes the plain formula
4	regression_2 errors cryptically on rank-deficient input	Low	Differs, but not better — see the doc
5	ena.rotation.h warns on every call	Trivial	n/a
6	ena.correlations(dims=) errors for any subset that is not 1:n	Low	Takes dimension names; any subset works

These are rENA issues, not ena-python bugs, reported as observations rather than accusations — they may be fixed upstream or intended. Only #1 changes ena-python's output relative to rENA:

Given both x_var and y_var, rotate_by_regression returns a different y axis than rENA, deliberately. rENA means to regress y on the x-deflated points (ena.rotate.by.regression.R sets V <- defA first), but that never takes effect: with.ena.matrix rebinds V to the raw points unless passed a V = argument, and a refactor dropped it. Its sibling ena.rotate.by.generalized deflates correctly on the same data. ena-python deflates, so its axes are orthogonal and its x axis matches rENA exactly. If rENA fixes this, ena-python will match it column-for-column.

The overwhelming majority of rENA matched ena-python exactly, kernel for kernel — and the two ena-python defects found in the same effort were worse (see CHANGELOG.md).

Development

python -m venv .venv && source .venv/bin/activate
python -m pip install -e ".[dev,plot,web]"
pytest

ruff check . && ruff format --check . && mypy src/ena_python && pytest

Regenerating fixtures needs R with rENA installed (install.packages("rENA")) — see reference/README.md. Everyday development does not need R.

Relationship to rENA, and citation

ena-python is an independent port. It is not affiliated with or endorsed by the rENA authors. All credit for the ENA method and the reference implementation belongs to the Epistemic Analytics group. If you use ena-python in research, cite the original ENA literature and rENA; see docs/migration_guide.md for the R-to-Python API map.

There is a separate, unrelated pyena package on PyPI by another author. This project is not it.

License

GPL-3.0-only, matching rENA (GPL-3), since ena-python is a derivative port. See LICENSE.