pycatdap 0.13.0

Python implementation of CATDAP (CATegorical Data Analysis Program)

pycatdap

AIC-based EDA and ML error analysis library for categorical data.

pycatdap is a Python implementation of CATDAP (CATegorical Data Analysis Program), developed by Sakamoto & Katsura (1980) at the Institute of Statistical Mathematics. It extends the classic CATDAP toolkit with modern exploratory data analysis (EDA) and machine learning error analysis workflows.

📖 Documentation: https://nbx-liz.github.io/pycatdap/

Why pycatdap?

Unlike general profilers (ydata-profiling, Skrub) or slice discovery tools (DivExplorer, pysubgroup), pycatdap uses AIC as its core relevance measure. This gives it four unique advantages:

Feature	Most tools	pycatdap
Variable relevance	Cramér's V, mutual info	AIC — explicit info-vs-complexity trade-off
Continuous binning	Equal-width or quantile	AIC-optimal binning
Subset discovery	Feature importance ranking	CATDAP-02 combinatorial search
Model coupling	Tied to specific frameworks	Model-agnostic (works with y_true, y_pred from anywhere)

Installation

# Core
pip install pycatdap

# With visualization (matplotlib)
pip install "pycatdap[plot]"

# With interactive Plotly figures + HTML reports
pip install "pycatdap[plotly]"

Supported: Python 3.10 / 3.11 / 3.12 / 3.13

Quickstart

Classic CATDAP

import pycatdap

df = pycatdap.datasets.load_health_data()

# CATDAP-01: pairwise AIC analysis
result = pycatdap.catdap1(df, response_names=["symptoms"])
print(result.aic_order["symptoms"])  # variables ranked by ΔAIC

# CATDAP-02: best explanatory subset
result2 = pycatdap.catdap2(
    df,
    pool=[2, 2, 2, 0, 0, 0, 0, 2],
    response_name="symptoms",
    accuracy=[0., 0., 0., 1., 1., 1., 0.1, 0.],
)
for s in result2.subsets[:3]:
    print(f"AIC={s.aic:.2f}, vars={s.variables}")

One-call EDA report (v0.5+)

report = pycatdap.profile(df, response="symptoms")
report.show()                       # Jupyter inline
report.to_html("report.html")       # self-contained HTML, inline Plotly
report.to_dict()                    # JSON-friendly
report.to_plotly_json()             # react-plotly.js / LizyStudio

ProfileResult exposes overview, variables (one VariableCard per column, including ΔAIC vs the response), association (m × m ΔAIC matrix), top_subsets (CATDAP-02 result when response is given), and quality_warnings (high_cardinality / constant / id_candidate / high_missing with overridable thresholds). See docs/tutorials/08-profile-titanic.ipynb for an end-to-end walkthrough.

Target analysis and CI-integrable suite (v0.6+)

# Target-driven: rank every column by ΔAIC vs `response`, keep top-K cross-tabs
ta = pycatdap.target_analysis(df, response="symptoms", top_k=5)
ta.ranking                         # variable / delta_aic / kind / n_obs
ta.top_summaries["cholesterol"]    # full TargetSummary for drill-down

# Quality scan only (fast — no catdap2 / association_matrix)
qr = pycatdap.quality_report(df)
assert qr.passed, qr.show()

# CI gate: one-line data contract
suite = pycatdap.suite.AICIndependenceSuite(df, response="symptoms")
result = suite.run()
assert result.passed, result.summary()

# Non-AIC association measures (pure-numpy, no scipy)
m = pycatdap.association_matrix(df, measure="cramers_v")
pycatdap.measures.register("my_measure", my_fn)  # pluggable per pysubgroup convention

See docs/tutorials/09-target-analysis-and-quality-suite.ipynb.

ML error labelling (v0.7.0)

from pycatdap.datasets import load_german_credit

df = load_german_credit()
y_true = df["class"]
y_pred = my_model.predict(df.drop(columns=["class"]))

# Label every prediction "correct" / "incorrect"
labels = pycatdap.error.error_label(y_true, y_pred)

# Or get fine-grained TP / FP / FN / TN (binary classification only)
conf = pycatdap.error.confusion_label(y_true, y_pred, positive="bad")

# Feed the error label back into target_analysis to discover which columns
# explain the errors — foundation for v0.8.0 `error_analysis()`
ta = pycatdap.target_analysis(df.assign(was_correct=labels), response="was_correct")

See docs/tutorials/10-ml-error-labeling.ipynb.

One-call ML error analysis (v0.8.0)

# task auto-detected from y_true / y_pred dtypes
result = pycatdap.error_analysis(
    df=test_df,
    y_true=y_test,                                  # str column name or array
    y_pred=model.predict(X_test),                   # str column name or array
    top_k=5,
)
result.show()                                       # Jupyter / stdout summary
result.to_html("errors.html")                       # self-contained Plotly report
result.feature_ranking                              # ΔAIC ranking of explanatories
result.top_slices                                   # single-variable cohort slices
result.confusion                                    # canonical TP/FP/FN/TN (binary)
result.to_divexplorer_format()                      # DivExplorer-compatible DataFrame

# Larger fairness-relevant benchmarks (needs `pip install 'pycatdap[data]'`)
df = pycatdap.datasets.fetch_california_housing()   # regression
df = pycatdap.datasets.fetch_adult_income()         # fairness demo
df = pycatdap.datasets.fetch_compas()               # fairness demo

See docs/tutorials/11-ml-error-analysis-one-call.ipynb.

ML error visualisation (v0.9.0)

r = pycatdap.error_analysis(df, y_true, y_pred)

# Result delegation — visualise straight off the result
r.plot_confusion()                                  # binary OR multi-class
r.plot_confusion(backend="plotly", normalize="true")
r.residual_plot()                                   # regression

# Or call the standalone functions
pycatdap.error.plot_confusion(y_true, y_pred, labels=[0, 1], normalize="true")
pycatdap.error.plot_confusion_by_slice(df, y_true, y_pred, var="age_group")
pycatdap.error.confusion_aic(y_true, y_pred)        # ΔAIC, negative = informative

pycatdap.error.residual_plot(y_true, y_pred, kind="histogram")
pycatdap.error.residual_by_category(df, y_true, y_pred, "feature")
pycatdap.error.residual_pool_plot(y_true, y_pred, n_bins=4)

See docs/tutorials/12-ml-error-visualization.ipynb.

ML calibration (v0.10.0)

r = pycatdap.error_analysis(df, y_true, y_pred, y_proba=proba)  # binary classification
r.calibration_curve(strategy="aic")                 # reliability diagram (AIC-binned)

# Or call the standalone functions
pycatdap.error.calibration_curve(y_true, proba, strategy="aic", backend="plotly")
pycatdap.error.brier_score(y_true, proba)
pycatdap.error.expected_calibration_error(y_true, proba)   # ECE
pycatdap.error.maximum_calibration_error(y_true, proba)    # MCE

strategy="aic" bins the probability axis where the observed positive-rate shifts — sharper than equal-width / quantile bins on skewed predictions. See docs/tutorials/13-ml-calibration.ipynb.

ML slice discovery, cohort comparison & drift (v0.11.0)

# Auto-discover the multivariable cohorts where the model fails most
result = pycatdap.error.discover_error_slices(
    df, y_true, y_pred, max_vars=3, measure="aic", top_k=10, min_support=30
)
for s in result.slices:
    print(s.error_metric, s.description)   # e.g. "age ∈ [60, 78] × plan = basic"
result.to_divexplorer_format()             # DivExplorer-compatible flat table

# Compare two cohorts (distribution + ΔAIC), Sweetviz-style HTML report
pycatdap.error.compare_cohorts(df_a, df_b).to_html("comparison.html")

# Detect train→prod drift, ranked by ΔAIC magnitude
pycatdap.error.detect_drift(df_train, df_prod, y_true=y, y_pred=yhat)

# Calibration beyond binary classification
pycatdap.error.regression_calibration_table(y_true, y_pred)          # regression
pycatdap.error.multiclass_calibration_table(y_true, y_proba)         # one-vs-rest

Slice discovery prunes the search space on support (Apriori, anti-monotone) rather than ΔAIC — sound, and >50% reduction on wide datasets. The interestingness measure is pluggable ("aic", "cramers_v", "mutual_info", or any registered callable). See docs/tutorials/14-ml-slice-discovery-and-drift.ipynb.

Status & Roadmap

Version	Theme
v0.2.0 ✅	Core CATDAP-01/02 (released)
v0.3.0 — v0.6.0 ✅	EDA workflow (Plotly backend, profile, target analysis)
v0.7.0 ✅	error labelling building blocks
v0.8.0 ✅	one-call error_analysis() + benchmark datasets
v0.9.0 ✅	error visualisation (confusion + residual)
v0.10.0 ✅	calibration (AIC-binned reliability diagram + Brier/ECE/MCE)
v0.11.0 ✅	slice discovery + cohort comparison + drift + regression/multi-class calibration
v0.12.0	LizyStudio integration
v1.0.0	API stabilization

Full roadmap: PLAN.md · Meta Issue #11

Development

git clone https://github.com/nbx-liz/pycatdap.git
cd pycatdap
uv sync --all-groups
uv run pytest                                  # tests (excluding slow R cross-validation)
uv run pytest -m slow                          # slow tests (requires R + catdap package)
uv run python -m mkdocs serve                  # local docs preview
make ci                                        # ruff + mypy + pytest + build

Contributing guidelines: CONTRIBUTING.md

Project structure

Document	Purpose
BLUEPRINT.md	Canonical specification (Japanese)
HISTORY.md	Proposal-to-decision log (Japanese)
PLAN.md	Development roadmap (Japanese)
CHANGELOG.md	Release history
docs/	Published documentation site

Citation

If you use pycatdap in research, please cite the original CATDAP work:

@article{sakamoto1980categorical,
  title={Categorical Data Analysis by AIC},
  author={Sakamoto, Yosiyuki and Katsura, Koichi},
  journal={Mathematical Sciences},
  year={1980}
}

License

MIT