Tutorials¶
Learning-oriented walkthroughs that guide you from zero to confident usage.
Available¶
- Basic CATDAP — The original CATDAP-01 / CATDAP-02 walkthrough on the UCI Heart Disease dataset
- EDA on Titanic — v0.3+ univariate EDA APIs (
describe,plot_variable,plot_missing) demonstrated on the Titanic dataset - AIC-optimal binning on iris — Why CATDAP-02's pooling search beats equal-width binning, with cut points visualized against species histograms
- Multivariate scale (synthetic) — Scale to 56 binary variables on a synthetic wide dataset, use
catdap1to rank candidates andcatdap2(nvar=k)to keep the subset search tractable - Real-world EDA on seaborn Titanic — End-to-end workflow on messy data: missingness inspection, cleanup, mixed-pooling CATDAP-02, and HTML/Plotly export. Requires
pycatdap[tutorial]. - Target-pair analysis — Drill into a single (target, explanatory) pair with
target_summaryandplot_target: four-perspective tables, AIC-optimal binning for continuous explanatories, ΔAIC ranking, Plotly export, and continuous-target regression mode viaRegressionTargetSummary(H-0005). Requirespycatdap[tutorial]. - Bivariate association analysis — Step beyond a single pair:
plot_pairfor symmetric two-column plots,aic_heatmapfor one-glance ΔAIC matrices,association_matrixfor the full m × m ΔAIC sweep, andassociation_plotfor vcd-style Pearson residuals. Requirespycatdap[tutorial]. - One-call profile() report — The v0.5.0 flagship API:
pycatdap.profile(df, response=...)bundles overview, variable cards, ΔAIC association matrix, CATDAP-02 top subsets, and quality warnings into a single call, then renders a self-contained HTML report with inline Plotly figures (H-0007). Requirespycatdap[plotly]andpycatdap[tutorial]. - Target analysis & quality suite — The v0.6.0 target-driven additions:
quality_report(df)for a fast CI-friendly quality scan,target_analysis(df, response)for ΔAIC ranking + top-K cross-tabs,pycatdap.measures.*(aic / cramers_v / mutual_info + register),association_matrix(measure=...)extension, andpycatdap.suite.AICIndependenceSuitefor theassert suite_result.passed, suite_result.summary()CI idiom. Requirespycatdap[plotly]andpycatdap[tutorial]. - ML error labelling — The v0.7.0 ML-error building blocks:
error_label,confusion_label,residual_label, andabs_residual_poolturn model predictions into CATDAP-ready categorical responses. - One-call ML error analysis — The v0.8.0 wrapper:
error_analysis(df, y_true, y_pred)returns anErrorAnalysisResultwith ΔAIC feature ranking, single-variable slices, and confusion/residual summaries. - ML error visualisation — The v0.9.0 visualisation layer:
plot_confusion/plot_confusion_by_slice/confusion_aic(classification) andresidual_plot/residual_by_category/residual_pool_plot(regression), plusErrorAnalysisResultdelegation methods on both backends. - ML calibration — The v0.10.0 calibration toolkit:
calibration_curvereliability diagram with AIC-optimal probability binning + Wilson CIs,brier_score/expected_calibration_error/maximum_calibration_error, and theErrorAnalysisResult.calibration_curvedelegation (binary classification). Requirespycatdap[plotly]andpycatdap[tutorial]. - ML slice discovery & drift — The v0.11.0 subgroup toolkit:
discover_error_slicesfinds multi-variable cohorts where model errors concentrate (support-pruned search, pluggable interestingness measure), pluscompare_cohortsanddetect_driftfor distribution/relationship shifts between datasets.