Expand description
§solow-metrics
Model-evaluation metrics for the Solow statistical stack — the numbers you report after a fit, not the ones an estimator uses internally.
The crate is organised by task:
regression— continuous-target losses and scores: mean squared error, RMSE, mean absolute error, MAPE, sMAPE, R², explained variance, max-error, mean pinball loss for quantile regression, Tweedie / Poisson / Gamma deviances, and their sample-weighted variants.classification— labels and probabilities: confusion matrix, accuracy, balanced accuracy, precision / recall / Fβ withmacro/micro/weighted/ per-class averaging, Matthews correlation, Cohen’s κ (linear and quadratic weighting), zero-one loss, hinge loss, binary log-loss, Brier score, and — for scored classifiers — the ROC and precision-recall curves, ROC-AUC (binary, one-vs-rest and Hand-Till one-vs-one for multiclass), and average precision.calibration— probability-calibration diagnostics: reliability curves with uniform or quantile bins, expected and maximum calibration error, and the Sanders / Murphy decomposition of the Brier score.forecast— time-series specific losses: MASE, RMSSE, pinball loss for prediction quantiles, interval coverage, the mean interval score (Winkler) for prediction bands, and the Harvey-Leybourne-Newbold small-sample-corrected Diebold-Mariano test for equal predictive accuracy of two forecasts.
All functions take ndarray views so they compose with slices, subranges, and
matrix rows without copies. Every metric returns a Result because pairs
with mismatched shapes, negative sample weights, unknown classes, or
degenerate inputs (a constant target for R², a single-class label vector for
ROC-AUC, …) are user errors, not silent NaNs.
The metric names, signatures, and averaging semantics deliberately mirror
the canonical the reference API so that ports and cross-checks are literal
— a reader can translate a the reference example one call at a time.
§Example
use ndarray::array;
use solow_metrics::{mean_squared_error, r2_score};
let y_true = array![3.0, -0.5, 2.0, 7.0];
let y_pred = array![2.5, 0.0, 2.0, 8.0];
let mse = mean_squared_error(y_true.view(), y_pred.view(), None).unwrap();
let r2 = r2_score(y_true.view(), y_pred.view(), None).unwrap();
assert!((mse - 0.375).abs() < 1e-12);
// R² = 1 - SS_res / SS_tot = 1 - 1.5 / 29.1875.
assert!((r2 - (1.0 - 1.5 / 29.1875)).abs() < 1e-10);Re-exports§
pub use bayesian::psis_loo;pub use bayesian::waic;pub use bayesian::PsisLooResult;pub use bayesian::WaicResult;pub use calibration::brier_decomposition;pub use calibration::expected_calibration_error;pub use calibration::maximum_calibration_error;pub use calibration::reliability_curve;pub use calibration::BinStrategy;pub use calibration::BrierDecomposition;pub use calibration::ReliabilityBin;pub use calibrators::IsotonicRegression;pub use calibrators::PlattScaling;pub use calibrators::TemperatureScaling;pub use classification::accuracy_score;pub use classification::average_precision_score;pub use classification::balanced_accuracy_score;pub use classification::binary_focal_loss;pub use classification::binary_log_loss;pub use classification::brier_score_loss;pub use classification::cohen_kappa_score;pub use classification::confusion_matrix;pub use classification::fbeta_score;pub use classification::hinge_loss;pub use classification::log_loss;pub use classification::matthews_corrcoef;pub use classification::multiclass_brier_score;pub use classification::multiclass_focal_loss;pub use classification::precision_recall_curve;pub use classification::precision_recall_fscore;pub use classification::ranked_probability_score;pub use classification::roc_auc_ovo_score;pub use classification::roc_auc_ovr_score;pub use classification::roc_auc_score;pub use classification::roc_curve;pub use classification::top_k_accuracy_score;pub use classification::top_label_calibration_error;pub use classification::zero_one_loss;pub use classification::Average;pub use classification::KappaWeights;pub use classification::MulticlassAuc;pub use classification::PrecisionRecallFScore;pub use classification::RocCurve as RocCurveResult;pub use comparison::friedman_test;pub use comparison::nemenyi_critical_difference;pub use comparison::wilcoxon_signed_rank;pub use comparison::FriedmanResult;pub use comparison::WilcoxonResult;pub use cluster::adjusted_mutual_info_score;pub use cluster::adjusted_rand_score;pub use cluster::calinski_harabasz_score;pub use cluster::completeness_score;pub use cluster::davies_bouldin_score;pub use cluster::fowlkes_mallows_score;pub use cluster::homogeneity_score;pub use cluster::normalized_mutual_info_score;pub use cluster::silhouette_score;pub use cluster::v_measure_score;pub use cluster::MiAverage;pub use conformal::JackknifePlus;pub use conformal::PredictionInterval;pub use conformal::SplitConformal;pub use effect_size::cliffs_delta;pub use effect_size::cohens_d;pub use effect_size::cramers_v;pub use effect_size::eta_squared;pub use effect_size::glass_delta;pub use effect_size::hedges_g;pub use effect_size::omega_squared;pub use forecast::diebold_mariano;pub use forecast::giacomini_white_test;pub use forecast::interval_coverage;pub use forecast::mase;pub use forecast::mean_interval_score;pub use forecast::pinball_loss;pub use forecast::rmsse;pub use forecast::DieboldMarianoResult;pub use forecast::DmLoss;pub use forecast::GiacominiWhiteResult;pub use inspection::accumulated_local_effects;pub use inspection::partial_dependence;pub use inspection::permutation_importance;pub use inspection::AccumulatedLocalEffects;pub use inspection::FeatureImportance;pub use inspection::PartialDependence;pub use pairwise::chi2_kernel;pub use pairwise::cosine_similarity;pub use pairwise::laplacian_kernel;pub use pairwise::linear_kernel;pub use pairwise::pairwise_distances;pub use pairwise::polynomial_kernel;pub use pairwise::rbf_kernel;pub use pairwise::sigmoid_kernel;pub use pairwise::PairwiseMetric;pub use report::classification_report;pub use report::ClassificationReport;pub use report::ClassificationRow;pub use regression::d2_absolute_error_score;pub use regression::d2_tweedie_score;pub use regression::explained_variance_score;pub use regression::huber_loss;pub use regression::log_cosh_loss;pub use regression::max_error;pub use regression::mean_absolute_error;pub use regression::mean_absolute_percentage_error;pub use regression::mean_gamma_deviance;pub use regression::mean_pinball_loss;pub use regression::mean_poisson_deviance;pub use regression::mean_squared_error;pub use regression::mean_squared_log_error;pub use regression::mean_tweedie_deviance;pub use regression::median_absolute_error;pub use regression::r2_score;pub use regression::root_mean_squared_error;pub use regression::root_mean_squared_log_error;pub use regression::symmetric_mean_absolute_percentage_error;pub use regression::RegressionReport;
Modules§
- bayesian
- Bayesian model comparison via WAIC and PSIS-LOO.
- calibration
- Probability-calibration diagnostics for binary probabilistic classifiers.
- calibrators
- Post-hoc probability calibrators.
- classification
- Classification metrics — confusion matrix, precision/recall/Fβ, log-loss, Brier score, ROC-AUC and average precision.
- cluster
- Cluster evaluation metrics — the reference
metrics.clustersurface. - comparison
- Statistical tests for comparing several models across benchmarks.
- conformal
- Distribution-free prediction intervals — conformal prediction.
- effect_
size - Effect-size metrics — a piece of solow’s world-class inference surface that has no direct the reference equivalent.
- forecast
- Forecasting-specific metrics.
- inspection
- Model-agnostic feature-importance diagnostics.
- pairwise
- Pairwise distances and kernels — the reference
metrics.pairwisesurface. - prelude
- Commonly used imports.
- regression
- Regression metrics — losses and scores for continuous targets.
- report
classification_report— the reference text summary of per-class precision, recall, F1, and support.