MLReview

MLReview provides model evaluation, inspection, diagnostics, and reporting for scikit-learn estimators. It replaces the extended-sklearn-metrics distribution with a smaller canonical API and a review-oriented result model.

Install

pip install ml-review

SHAP inspection is optional:

pip install ml-review[shap]

MLReview uses three spellings deliberately:

Quick Start

from ml_review import evaluate
from sklearn.datasets import make_classification
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split

X, y = make_classification(n_samples=500, n_features=8, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)

model = RandomForestClassifier(random_state=42).fit(X_train, y_train)
review = evaluate(model, X_train, y_train, X_test, y_test)

summary = review.report()
importance = review.importance.report()

evaluate(...) returns ReviewResult. User-facing reports are DataFrames, while rich result state remains available through mapping reads and raw export:

review["performance"]
review.get("feature_importance")
raw = review.to_dict()

Canonical API

Keep the top-level import small:

from ml_review import evaluate
from ml_review.evaluation import Thresholds, classification_cv, regression_cv

Use namespace imports for grouped workflows:

Result Workflows

review.report()
review.print_summary()

review.importance.report()
review.importance.plot()

review.fairness.report()
review.fairness.plot()

Methods are thin wrappers around shared functions, so the functional form also works:

from ml_review.inspection import importance
from ml_review.reporting import evaluation

evaluation.report(review)
importance.report(review)

Cross-validation summary APIs still return DataFrames directly:

from ml_review.evaluation import classification_cv, regression_cv

classification_table = classification_cv(classifier, X, y, cv=5)
regression_table = regression_cv(regressor, X, y, cv=5)

Optional SHAP Inspection

SHAP is preferred by evaluate(...) when it is installed. If it is missing in the default auto mode, MLReview keeps built-in and permutation importance behavior. Use shap_mode="on" to require SHAP or shap_mode="off" to skip it.

review = evaluate(
    model,
    X_train,
    y_train,
    X_test,
    y_test,
    shap_mode="auto",
    shap_background_size=100,
    shap_sample_size=200,
)

from ml_review.inspection import shap

shap_table = review.shap.report()
review.shap.plot_importance()
review.shap.plot_explanation(sample_index=0)

# Equivalent functional calls
shap.report(review)
shap.plot_importance(review)

SHAP guidance in the report:

• the baseline is the model output before feature contributions are added
• positive SHAP values raise the explained output and negative values lower it
• global SHAP importance averages absolute contributions across sampled rows
• local SHAP plots explain one stored sampled prediction
• classifier local plots default to the predicted class unless an output index is provided

MLReview stores bounded serialized SHAP payloads in review data rather than native SHAP explanation objects.

ROC and Residual Review

ROC functions receive an estimator and data, matching the existing scikit-learn-oriented implementation:

from ml_review.metrics import roc

roc_result = roc.binary(classifier, X, y, cv=5)
roc_table = roc_result.report()
thresholds = roc_result.thresholds()

pr_result = roc.precision_recall(classifier, X, y, cv=5)
pr_table = pr_result.report()

Residual diagnostics also use a result object with DataFrame reports:

from ml_review.diagnostics import residual

residual_result = residual.calculate(regressor, X, y, cv=5)
residual_table = residual_result.report()
residual_result.plot()

Migration

extended_sklearn_metrics remains importable from the new distribution as a warning-bearing compatibility path. New releases are published as ml-review.

See MIGRATION.md for import examples and compatibility notes.

Roadmap

The MLReview foundation release covers the package rename, namespace cleanup, result objects, DataFrame-first reports, compatibility shims, and optional SHAP inspection. Planned later feature releases cover:

1. calibration and probability reliability review
2. unsupervised review for clustering, PCA, and embeddings
3. deeper inspection, fairness checks, and report artifacts

See ROADMAP.md for the feature pipeline and guardrails.

License

MIT