Migrating from scikit-plot¶

scikit-plot (skplt) provides one-function plotting for sklearn model evaluation — plot_roc, plot_confusion_matrix, plot_precision_recall, etc. Ferrum's *_chart helpers follow the same pattern: pass a fitted model and held-out data, get a chart back.

Function mapping¶

scikit-plot	Ferrum	Notes
`skplt.metrics.plot_roc(y_test, y_probas)`	`fm.roc_chart(model, X_test, y_test)`	Ferrum takes the model directly; computes probabilities internally.
`skplt.metrics.plot_precision_recall(y_test, y_probas)`	`fm.pr_chart(model, X_test, y_test)`	Same.
`skplt.metrics.plot_confusion_matrix(y_test, y_pred)`	`fm.confusion_matrix_chart(model, X_test, y_test)`	`normalize=` supported.
`skplt.metrics.plot_calibration_curve(y_test, probas_list)`	`fm.calibration_chart(model, X_test, y_test)`	Single model or `ComparedModelSource` for multi-model.
`skplt.metrics.plot_cumulative_gain(y_test, y_probas)`	`fm.gain_chart(model, X_test, y_test)`	—
`skplt.metrics.plot_lift_curve(y_test, y_probas)`	`fm.lift_chart(model, X_test, y_test)`	—
`skplt.estimators.plot_feature_importances(model)`	`fm.importance_chart(model, X_test, y_test)`	Handles `feature_importances_`, permutation, and coefficients.
`skplt.estimators.plot_learning_curve(model, X, y)`	`fm.learning_curve_chart(model, X, y)`	—
`skplt.cluster.plot_silhouette(X, labels)`	`fm.silhouette_chart(model, X)`	Also available as `SilhouetteVisualizer` for lifecycle control.
`skplt.cluster.plot_elbow_curve(model, X)`	`fm.elbow_chart(model, X)`	Also available as `ElbowVisualizer` for lifecycle control.
`skplt.decomposition.plot_pca_component_variance(pca)`	`fm.pca_scree_chart(model, X)`	—

Key differences¶

Model-in, chart-out¶

scikit-plot typically requires you to pre-compute predictions (y_pred) or probabilities (y_probas) and pass arrays. Ferrum takes the fitted model directly and computes everything internally via ModelSource:

# scikit-plot
y_probas = model.predict_proba(X_test)
skplt.metrics.plot_roc(y_test, y_probas)
plt.savefig("roc.svg")

# Ferrum
chart = fm.roc_chart(model, X_test, y_test)
chart.save("roc.svg")

No matplotlib¶

scikit-plot renders to matplotlib and requires plt.show() or plt.savefig(). Ferrum returns a Chart — same object you get from fm.Chart(data).mark_point().encode(...).

Charts compose¶

scikit-plot functions are standalone — combining multiple plots means managing matplotlib subplots. In Ferrum, diagnostic charts compose with the same operators as any other chart:

roc = fm.roc_chart(model, X_test, y_test)
cm = fm.confusion_matrix_chart(model, X_test, y_test)
gain = fm.gain_chart(model, X_test, y_test)
report = (roc | cm) & gain
report.save("model_report.svg")

ModelSource for efficiency¶

When running multiple diagnostics, build a ModelSource once:

source = fm.ModelSource(model, X_test, y_test)
roc = fm.roc_chart(source)
cm = fm.confusion_matrix_chart(source)
pr = fm.pr_chart(source)

Predicted probabilities, confusion counts, and other derived tables are computed once and shared.

Themes¶

scikit-plot inherits matplotlib styling. Ferrum charts accept theme values:

fm.roc_chart(model, X_test, y_test).theme(fm.themes.publication).save("roc.svg")

Coverage comparison¶

Ferrum's diagnostic surface is a superset of scikit-plot's:

Category	scikit-plot	Ferrum
Classification metrics	ROC, PR, confusion matrix, calibration, gain, lift	All of scikit-plot's + class prediction error, discrimination threshold
Regression	—	Residuals, prediction error, Cook's distance
Feature explanation	Feature importances	Importances + SHAP (beeswarm, bar, waterfall) + partial dependence
Model selection	Learning curve	Learning curve + validation curve + CV scores + alpha selection
Clustering	Silhouette, elbow, PCA variance	All of scikit-plot's + intercluster distance, decision boundary, parallel coordinates

Where to go next¶

Model diagnostics for the full diagnostic surface.
Figure-level helpers for the one-line helper pattern.
Composition for building multi-panel model reports.