FAQ

Why did screening find no rules? Every candidate must clear the support floor on both sides — by default min_support = min(200, max(20, ceil(0.02 * n_train))) — and survive BH-FDR at fdr_alpha (default 0.05). On small datasets or weak signals nothing may qualify; a feature is skipped outright if it has fewer than 2 * min_support non-missing values. Pass a smaller explicit min_support, raise fdr_alpha, or bring more rows. See rules and screening.

Why does export_rules() / explain() raise ValueError? You fitted with representation="compact". The compact head's weights are per-class scores over collapsed flag counts, not per-rule coefficients, so there is no rule table to export. Refit with representation="full" (the default) to get rules and reason codes.

How are missing values handled — do I need to impute? No. By default (missing="no_evidence") a missing value never triggers a flag and contributes nothing to the score. With missing="indicator", missingness itself is screened like any other candidate and can become a missing_indicator rule. See missing values.

Why does importing the plots fail with an ImportError? The matplotlib helpers live behind the optional viz extra: pip install "flaggam[viz]". Importing flaggam itself never requires matplotlib — the error is raised lazily, only when a plotting function is called.

Why does export_rules_html refuse my estimator? The explorer needs per-rule weights, so it requires representation="full" and the additive head, and supports binary classification and regression only (a multiclass head has one coefficient vector per class, so there is no single shape per feature). The same limits apply to plot_shape.

Can I use a tree ensemble on top of the rules? Yes — pass any scikit-learn estimator as head="flexible". It is fit on the rule basis Z(X) with no access to the raw features, but you trade away additive interpretability: export_rules() and explain() mark the result additive_interpretable=False with weight=NaN.

Is the monotonicity constraint exact or approximate? Exact. A feature's bases are themselves monotone step/ramp functions, so sign-constraining their coefficients makes the fitted shape monotone at every value — no post-hoc projection. Supported for representation="full" binary classification and regression; a list-valued C/alpha falls back to 1.0. See monotonicity.

Can I put CalibratedFlagGAM inside a Pipeline or GridSearchCV? Not currently — it is a thin wrapper, not a full scikit-learn estimator, so it does not support clone(). Tune the underlying FlagGAMClassifier first, then calibrate the chosen configuration at the top level. The base estimators themselves pass check_estimator and work anywhere scikit-learn estimators do.

How close does this implementation get to the paper's numbers? The German Credit smoke benchmark lands at 0.773 AUROC against the paper's reported 0.775, and the runners reproduce the protocol of Tables 3, 4, 5, 7, and 8 with --n-splits 1000 defaults matching the paper. See benchmarks.