Visualization

Interactive rules explorer

export_rules_html renders a fitted estimator's rules as a single dependency-free HTML file: no network access, no external CSS/JS, safe to open offline or embed in an iframe. It shows a feature-selector dropdown, the fitted shape curve (or level bars for categorical features), and a table of the rules contributing to that feature. The embed below is generated from a FlagGAMClassifier fit on the German Credit dataset — regenerate it with uv run python scripts/make_rules_explorer.py.

from flaggam import export_rules_html

export_rules_html(clf, path="rules.html")  # open in any browser

flaggam.plots provides matplotlib helpers for fitted estimators and diagnostics. This module requires the optional viz extra:

uv sync --extra viz

Importing flaggam never imports matplotlib eagerly — each plotting function calls a lazy import helper first, raising a clear ImportError naming pip install flaggam[viz] if matplotlib is not installed. Every function accepts an optional ax and returns the Axes it drew on, so plots compose with existing matplotlib figures.

Shape Function — plot_shape

Plots the fitted additive contribution for one feature: a step curve of contribution vs. value (with a rug of the discovered cutoffs) for numeric features, or one bar per discovered level for categorical features. Requires representation="full".

from flaggam import plot_shape

ax = plot_shape(clf, "age")

Rule Importance — plot_rule_importance

Horizontal bar chart of the top-N rules by |weight|, read directly from export_rules().

from flaggam import plot_rule_importance

ax = plot_rule_importance(clf, top_n=20)

Waterfall — plot_waterfall

Cumulative horizontal bars from the intercept to the total score for a single row, sourced from explain(x_row). Rules beyond max_rules collapse into a single "(other rules)" bucket so the chart stays readable.

from flaggam import plot_waterfall

x_row = X.iloc[[0]]
ax = plot_waterfall(clf, x_row, max_rules=15)

Reliability Diagram — plot_reliability

Mean predicted probability vs. observed positive fraction per bin, with a per-bin count overlay, built on reliability_curve from flaggam.calibration.

from flaggam import plot_reliability

ax = plot_reliability(y_test, clf.predict_proba(X_test)[:, 1], n_bins=10)

Proxy Association — plot_proxy_association

Horizontal bars of rule-level association with a protected attribute, from a ProxyAudit(...).report(...) DataFrame; bars whose association exceeds the audit threshold are highlighted, with a dashed line marking the approximate cutoff.

from flaggam import plot_proxy_association, ProxyAudit

report = ProxyAudit(clf).report(X, A)
ax = plot_proxy_association(report, top_n=20)

Group Metrics — plot_group_metrics

Grouped bar chart of selection_rate, tpr, and auroc per level of a protected attribute, from a group_metrics(...) result; the title annotates the computed fairness gaps.

from flaggam import plot_group_metrics, group_metrics

metrics = group_metrics(y_test, clf.predict_proba(X_test)[:, 1], A_test)
ax = plot_group_metrics(metrics)

See Plots for the full API reference.