ferrolearn-linear 0.5.0

Linear models for the ferrolearn ML framework
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
//! # ferrolearn-linear
//!
//! Linear models for the ferrolearn machine learning framework.
//!
//! This crate provides implementations of the most common linear models
//! for both regression and classification tasks:
//!
//! - **[`LinearRegression`]** — Ordinary Least Squares via QR decomposition
//! - **[`Ridge`]** — L2-regularized regression via Cholesky decomposition
//! - **[`RidgeCV`]** — Ridge with built-in cross-validated alpha selection
//! - **[`Lasso`]** — L1-regularized regression via coordinate descent
//! - **[`LassoCV`]** — Lasso with built-in cross-validated alpha selection
//! - **[`ElasticNet`]** — Combined L1/L2 regularization via coordinate descent
//! - **[`ElasticNetCV`]** — ElasticNet with cross-validated (alpha, l1_ratio) selection
//! - **[`BayesianRidge`]** — Bayesian Ridge with automatic regularization tuning
//! - **[`HuberRegressor`]** — Robust regression via IRLS with Huber loss
//! - **[`LogisticRegression`]** — Binary and multiclass classification via L-BFGS
//!
//! All models implement the [`ferrolearn_core::Fit`] and [`ferrolearn_core::Predict`]
//! traits, and produce fitted types that implement [`ferrolearn_core::introspection::HasCoefficients`].
//!
//! # Design
//!
//! Each model follows the compile-time safety pattern:
//!
//! - The unfitted struct (e.g., `LinearRegression<F>`) holds hyperparameters
//!   and implements [`Fit`](ferrolearn_core::Fit).
//! - Calling `fit()` produces a new fitted type (e.g., `FittedLinearRegression<F>`)
//!   that implements [`ferrolearn_core::Predict`].
//! - Calling `predict()` on an unfitted model is a compile-time error.
//!
//! # Pipeline Integration
//!
//! All models implement [`PipelineEstimator`](ferrolearn_core::pipeline::PipelineEstimator)
//! for `f64`, allowing them to be used as the final step in a
//! [`Pipeline`](ferrolearn_core::pipeline::Pipeline).
//!
//! # Float Generics
//!
//! All models are generic over `F: num_traits::Float + Send + Sync + 'static`,
//! supporting both `f32` and `f64`.
//!
//! # `## REQ status`
//!
//! Binary (R-DEFER-2) for the crate-root RE-EXPORT BOUNDARY (this file is the public-API
//! surface, not an estimator). Mirrors `sklearn/linear_model/__init__.py` `__all__`
//! (`:48-98`) + the `score` mixins `sklearn/base.py` `ClassifierMixin.score` (`:738-764`,
//! accuracy) / `RegressorMixin.score` (`:805-849`, R²). Design doc: `.design/linear/lib.md`.
//! Per-estimator REQs live in the sibling modules' own routed docs. Score traits
//! (`ClassifierScore`/`RegressorScore`) are pre-existing crate-root `pub trait`s re-exported
//! via the meta-crate (`ferrolearn::linear`) — grandfathered public API (goal.md S5); honest
//! underclaim (R-HONEST-3): no production `.score()` caller yet. `sample_weight` is
//! supported via the `Option<&Array1<F>>` last param on both traits (REQ-6, #1106).
//!
//! | REQ | Status | Evidence |
//! |---|---|---|
//! | REQ-1 (re-export boundary) | SHIPPED | the `pub use` block re-exports every implemented linear/svm/discriminant_analysis/isotonic estimator at the crate root, mirroring sklearn `linear_model.__all__` (`__init__.py:48-98`), broadened per goal.md scope §2. Consumers: meta-crate `pub use ferrolearn_linear as linear` + PyO3 shim `ferrolearn-python/src/{regressors,classifiers,extras}.rs`. |
//! | REQ-2 (`ClassifierScore::score` == mean accuracy) | SHIPPED | `ClassifierScore` blanket impl body `mean_accuracy` (`correct / n`) mirrors `ClassifierMixin.score` → `accuracy_score` (`base.py:738-764`); critic-verified vs live oracle (`accuracy_score([0,1,2,1],[0,1,1,1])=0.75`). Consumer: grandfathered crate/meta re-export (S5). Underclaim: no production `.score()` caller; single-label (`Output=Array1<usize>`). |
//! | REQ-3 (`RegressorScore::score` == in-regime R²) | SHIPPED | `RegressorScore` blanket impl body `r2_score` = `1 − ss_res/ss_tot` mirrors `RegressorMixin.score` → `metrics.r2_score` (`base.py:805-849`); matches live oracle `r2_score([3.,5.,2.,7.],[2.5,5.,2.,8.])=0.9152542372881356` (`r2_in_regime_matches_oracle`). Consumer: grandfathered re-export (S5). |
//! | REQ-4 (constant-y R² edge parity) | SHIPPED | FIXED #1104. `r2_score` now returns `0.0` (was `neg_infinity()`) when `ss_tot==0 ∧ ss_res!=0`, matching `metrics.r2_score` (`_regression.py:891`); zero-residual stays `1.0`. Green: `divergence_r2_constant_ytrue_nonzero_residual_returns_zero` + `r2_constant_ytrue_zero_residual_returns_one`. |
//! | REQ-5 (`log_proba` behind predict_log_proba) | SHIPPED | FIXED #1105. `log_proba` is now the unclamped `p.ln()`, matching sklearn `predict_log_proba = np.log(predict_proba)` (`discriminant_analysis.py:1059`); `0.0`→`-inf`. Consumers: `logistic_regression.rs`/`logistic_regression_cv.rs`/`qda.rs` `predict_log_proba`. Green: `divergence_log_proba_zero_clamps_instead_of_neg_inf`. |
//! | REQ-6 (sample_weight on score) | SHIPPED | FIXED #1106. Both score traits now take `sample_weight: Option<&Array1<F>>` as the LAST param, matching sklearn `score(self, X, y, sample_weight=None)` (`base.py:738`,`:805`). `ClassifierScore::score` → `weighted_accuracy` (`Σ wᵢ·[predᵢ==yᵢ] / Σ wᵢ`, the `accuracy_score(..., sample_weight=w)` analog); `RegressorScore::score` → `weighted_r2_score` (`1 − Σwᵢ(yᵢ−predᵢ)² / Σwᵢ(yᵢ−ȳ_w)²`, ȳ_w = `Σwᵢyᵢ/Σwᵢ`, the `r2_score(..., sample_weight=w)` analog). `None` is byte-identical to `mean_accuracy`/`r2_score`. Oracle-verified (live sklearn 1.5.2): `weighted_accuracy([0,1,0,0,1],[0,1,1,0,1],[1,2,3,1,1])=0.625`, `weighted_r2_score([1.1,1.9,3.2,3.7,5.1],[1,2,3,4,5],[1,2,3,1,1])=0.9770114942528736` in `tests/divergence_lib.rs` (`classifier_score_weighted_matches_sklearn`, `regressor_score_weighted_matches_sklearn`, `score_none_sample_weight_equals_unweighted`). |
//! | REQ-substrate (ferray) | NOT-STARTED | open prereq blocker #1107. Helpers + score traits run on `ndarray::{Array1,Array2}` + `num_traits::Float`, not `ferray-core` arrays (R-SUBSTRATE-1). |

pub mod ard;
pub mod bayesian_ridge;
pub mod elastic_net;
pub mod elastic_net_cv;
pub mod glm;
pub mod huber_regressor;
pub mod isotonic;
pub mod lars;
pub mod lasso;
pub mod lasso_cv;
pub mod lda;
mod linalg;
pub mod linear_regression;
pub mod linear_svc;
pub mod linear_svr;
pub mod logistic_regression;
pub mod logistic_regression_cv;
pub mod multi_task_elastic_net;
pub mod multi_task_elastic_net_cv;
pub mod multi_task_lasso;
pub mod multi_task_lasso_cv;
pub mod nu_svm;
pub mod omp;
pub mod one_class_svm;
mod optim;
pub mod qda;
pub mod quantile_regressor;
pub mod ransac;
pub mod ridge;
pub mod ridge_classifier;
pub mod ridge_classifier_cv;
pub mod ridge_cv;
pub mod sgd;
pub mod svm;

// Re-export the main types at the crate root.
pub use ard::{ARDRegression, FittedARDRegression};
pub use bayesian_ridge::{BayesianRidge, FittedBayesianRidge};
pub use elastic_net::{ElasticNet, FittedElasticNet};
pub use elastic_net_cv::{ElasticNetCV, FittedElasticNetCV};
pub use glm::{
    FittedGLMRegressor, GLMFamily, GLMRegressor, GammaRegressor, PoissonRegressor, TweedieRegressor,
};
pub use huber_regressor::{FittedHuberRegressor, HuberRegressor};
pub use isotonic::{FittedIsotonicRegression, IsotonicRegression};
pub use lars::{FittedLars, FittedLassoLars, Lars, LassoLars};
pub use lasso::{FittedLasso, Lasso};
pub use lasso_cv::{FittedLassoCV, LassoCV};
pub use lda::{FittedLDA, LDA};
pub use linear_regression::{FittedLinearRegression, LinearRegression};
pub use linear_svc::{FittedLinearSVC, LinearSVC, LinearSVCLoss};
pub use linear_svr::{FittedLinearSVR, LinearSVR, LinearSVRLoss};
pub use logistic_regression::{FittedLogisticRegression, LogisticRegression};
pub use logistic_regression_cv::{FittedLogisticRegressionCV, LogisticRegressionCV};
pub use multi_task_elastic_net::{FittedMultiTaskElasticNet, MultiTaskElasticNet};
pub use multi_task_elastic_net_cv::{FittedMultiTaskElasticNetCV, MultiTaskElasticNetCV};
pub use multi_task_lasso::{FittedMultiTaskLasso, MultiTaskLasso};
pub use multi_task_lasso_cv::{FittedMultiTaskLassoCV, MultiTaskLassoCV};
pub use nu_svm::{FittedNuSVC, FittedNuSVR, NuSVC, NuSVR};
pub use omp::{FittedOMP, OrthogonalMatchingPursuit};
pub use one_class_svm::{FittedOneClassSVM, OneClassSVM};
pub use qda::{FittedQDA, QDA};
pub use quantile_regressor::{FittedQuantileRegressor, QuantileRegressor};
pub use ransac::{FittedRANSACRegressor, MinSamples, RANSACRegressor, RansacLoss};
pub use ridge::{FittedRidge, FittedRidgeMulti, Ridge};
pub use ridge_classifier::{FittedRidgeClassifier, RidgeClassifier};
pub use ridge_classifier_cv::{FittedRidgeClassifierCV, RidgeClassifierCV};
pub use ridge_cv::{FittedRidgeCV, RidgeCV};
pub use sgd::{
    FittedSGDClassifier, FittedSGDOneClassSVM, FittedSGDRegressor, SGDClassifier, SGDOneClassSVM,
    SGDRegressor,
};
pub use svm::{
    FittedSVC, FittedSVR, Kernel, LinearKernel, PolynomialKernel, RbfKernel, SVC, SVR,
    SigmoidKernel,
};

use ferrolearn_core::error::FerroError;
use ferrolearn_core::traits::Predict;
use ndarray::{Array1, Array2};
use num_traits::Float;

/// Mean-accuracy `score(x, y)` exposed on every fitted classifier in this
/// crate via a blanket impl over [`Predict<Array2<F>, Output=Array1<usize>>`].
///
/// Users just `use ferrolearn_linear::ClassifierScore;` to call
/// `fitted.score(&x, &y)` and get the same result as sklearn's
/// `ClassifierMixin.score`.
pub trait ClassifierScore<F: Float> {
    /// (Optionally weighted) mean accuracy on the given test data and labels.
    ///
    /// Mirrors sklearn `ClassifierMixin.score(self, X, y, sample_weight=None)`
    /// (`base.py:738`) → `accuracy_score(y, predict(X), sample_weight=...)`.
    /// Passing `None` for `sample_weight` reproduces the unweighted
    /// `correct / n` accuracy.
    ///
    /// # Errors
    ///
    /// Returns [`FerroError::ShapeMismatch`] if `x.nrows() != y.len()`, or if a
    /// non-`None` `sample_weight` has a length other than `y.len()`, or any
    /// error forwarded from the inner `predict`.
    fn score(
        &self,
        x: &Array2<F>,
        y: &Array1<usize>,
        sample_weight: Option<&Array1<F>>,
    ) -> Result<F, FerroError>;
}

impl<T, F> ClassifierScore<F> for T
where
    T: Predict<Array2<F>, Output = Array1<usize>, Error = FerroError>,
    F: Float,
{
    fn score(
        &self,
        x: &Array2<F>,
        y: &Array1<usize>,
        sample_weight: Option<&Array1<F>>,
    ) -> Result<F, FerroError> {
        if x.nrows() != y.len() {
            return Err(FerroError::ShapeMismatch {
                expected: vec![x.nrows()],
                actual: vec![y.len()],
                context: "y length must match number of samples in X".into(),
            });
        }
        let preds = self.predict(x)?;
        weighted_accuracy(&preds, y, sample_weight)
    }
}

/// R² `score(x, y)` exposed on every fitted regressor in this crate via
/// a blanket impl over [`Predict<Array2<F>, Output=Array1<F>>`].
///
/// Users just `use ferrolearn_linear::RegressorScore;` to call
/// `fitted.score(&x, &y)`.
pub trait RegressorScore<F: Float> {
    /// (Optionally weighted) R² coefficient of determination on the given test
    /// data and targets.
    ///
    /// Mirrors sklearn `RegressorMixin.score(self, X, y, sample_weight=None)`
    /// (`base.py:805`) → `r2_score(y, predict(X), sample_weight=...)`. Passing
    /// `None` for `sample_weight` reproduces the unweighted `1 − SSres/SStot`.
    ///
    /// # Errors
    ///
    /// Returns [`FerroError::ShapeMismatch`] if `x.nrows() != y.len()`, or if a
    /// non-`None` `sample_weight` has a length other than `y.len()`, or any
    /// error forwarded from the inner `predict`.
    fn score(
        &self,
        x: &Array2<F>,
        y: &Array1<F>,
        sample_weight: Option<&Array1<F>>,
    ) -> Result<F, FerroError>;
}

impl<T, F> RegressorScore<F> for T
where
    T: Predict<Array2<F>, Output = Array1<F>, Error = FerroError>,
    F: Float,
{
    fn score(
        &self,
        x: &Array2<F>,
        y: &Array1<F>,
        sample_weight: Option<&Array1<F>>,
    ) -> Result<F, FerroError> {
        if x.nrows() != y.len() {
            return Err(FerroError::ShapeMismatch {
                expected: vec![x.nrows()],
                actual: vec![y.len()],
                context: "y length must match number of samples in X".into(),
            });
        }
        let preds = self.predict(x)?;
        weighted_r2_score(&preds, y, sample_weight)
    }
}

/// Mean accuracy: `(sum(predictions == targets)) / n`.
///
/// Used as the body of every classifier `score(&self, x, y)` method in
/// this crate to mirror sklearn's `ClassifierMixin.score`.
pub(crate) fn mean_accuracy<F: Float>(predictions: &Array1<usize>, targets: &Array1<usize>) -> F {
    let n = targets.len();
    if n == 0 {
        return F::zero();
    }
    let correct = predictions
        .iter()
        .zip(targets.iter())
        .filter(|(p, t)| p == t)
        .count();
    F::from(correct).map_or(F::zero(), |c| c / F::from(n).unwrap_or(F::one()))
}

/// (Optionally) weighted mean accuracy:
/// `sum_i(w_i · [pred_i == y_i]) / sum_i(w_i)`.
///
/// Mirrors sklearn `ClassifierMixin.score` → `accuracy_score(y, pred,
/// sample_weight=...)` (`base.py:738-764`; `metrics._classification.accuracy_score`
/// forwards `sample_weight` into `_weighted_sum`). When `sample_weight` is
/// `None`, the result is byte-identical to [`mean_accuracy`] (`correct / n`).
///
/// # Errors
///
/// Returns [`FerroError::ShapeMismatch`] if a non-`None` `sample_weight` has a
/// length other than `targets.len()`.
pub(crate) fn weighted_accuracy<F: Float>(
    predictions: &Array1<usize>,
    targets: &Array1<usize>,
    sample_weight: Option<&Array1<F>>,
) -> Result<F, FerroError> {
    let Some(w) = sample_weight else {
        return Ok(mean_accuracy(predictions, targets));
    };
    let n = targets.len();
    if w.len() != n {
        return Err(FerroError::ShapeMismatch {
            expected: vec![n],
            actual: vec![w.len()],
            context: "sample_weight length must match number of samples".into(),
        });
    }
    if n == 0 {
        return Ok(F::zero());
    }
    let mut num = F::zero();
    let mut den = F::zero();
    for i in 0..n {
        den = den + w[i];
        if predictions[i] == targets[i] {
            num = num + w[i];
        }
    }
    Ok(num / den)
}

/// R² coefficient of determination: `1 - SSres / SStot`. Used as the
/// body of every regressor `score(&self, x, y)` method to mirror
/// sklearn's `RegressorMixin.score`. Constant-y returns `1.0` if
/// predictions are also constant-perfect (zero residual), else `0.0`
/// when the residual is non-zero — matching `sklearn.metrics.r2_score`
/// (`_regression.py:891`: `output_scores[nonzero_numerator &
/// ~nonzero_denominator] = 0.0`).
pub(crate) fn r2_score<F: Float>(y_pred: &Array1<F>, y_true: &Array1<F>) -> F {
    let n = y_true.len();
    if n == 0 {
        return F::zero();
    }
    let mean = y_true.iter().copied().fold(F::zero(), |a, b| a + b) / F::from(n).unwrap();
    let mut ss_res = F::zero();
    let mut ss_tot = F::zero();
    for i in 0..n {
        let r = y_true[i] - y_pred[i];
        let t = y_true[i] - mean;
        ss_res = ss_res + r * r;
        ss_tot = ss_tot + t * t;
    }
    if ss_tot == F::zero() {
        if ss_res == F::zero() {
            F::one()
        } else {
            F::zero()
        }
    } else {
        F::one() - ss_res / ss_tot
    }
}

/// (Optionally) weighted R² coefficient of determination:
/// `1 − SSres_w / SStot_w`, where `SSres_w = Σ w_i·(y_i − pred_i)²`,
/// `SStot_w = Σ w_i·(y_i − ȳ_w)²`, and `ȳ_w = Σ w_i·y_i / Σ w_i`.
///
/// Mirrors sklearn `RegressorMixin.score` → `r2_score(y, pred,
/// sample_weight=...)` (`base.py:805-849`; `metrics._regression.r2_score`
/// weights both numerator and denominator by `sample_weight`). When
/// `sample_weight` is `None`, the result is byte-identical to [`r2_score`].
/// The constant-`y` edge (`SStot_w == 0`) matches the unweighted convention:
/// `1.0` if the (weighted) residual is also `0`, else `0.0`.
///
/// # Errors
///
/// Returns [`FerroError::ShapeMismatch`] if a non-`None` `sample_weight` has a
/// length other than `y_true.len()`.
pub(crate) fn weighted_r2_score<F: Float>(
    y_pred: &Array1<F>,
    y_true: &Array1<F>,
    sample_weight: Option<&Array1<F>>,
) -> Result<F, FerroError> {
    let Some(w) = sample_weight else {
        return Ok(r2_score(y_pred, y_true));
    };
    let n = y_true.len();
    if w.len() != n {
        return Err(FerroError::ShapeMismatch {
            expected: vec![n],
            actual: vec![w.len()],
            context: "sample_weight length must match number of samples".into(),
        });
    }
    if n == 0 {
        return Ok(F::zero());
    }
    let mut w_sum = F::zero();
    let mut wy_sum = F::zero();
    for i in 0..n {
        w_sum = w_sum + w[i];
        wy_sum = wy_sum + w[i] * y_true[i];
    }
    let mean = wy_sum / w_sum;
    let mut ss_res = F::zero();
    let mut ss_tot = F::zero();
    for i in 0..n {
        let r = y_true[i] - y_pred[i];
        let t = y_true[i] - mean;
        ss_res = ss_res + w[i] * r * r;
        ss_tot = ss_tot + w[i] * t * t;
    }
    if ss_tot == F::zero() {
        if ss_res == F::zero() {
            Ok(F::one())
        } else {
            Ok(F::zero())
        }
    } else {
        Ok(F::one() - ss_res / ss_tot)
    }
}

/// Element-wise natural log of a probability matrix, used as the body of every
/// classifier `predict_log_proba` method in this crate. Unclamped, mirroring
/// scikit-learn `predict_log_proba = np.log(predict_proba)`
/// (`sklearn/discriminant_analysis.py:1059`: `return np.log(probas_)`): a `0.0`
/// probability maps to `-inf`. Inputs are always in `[0, 1]`, so the result is
/// either finite or `-inf` (never `NaN`).
pub(crate) fn log_proba<F: Float>(proba: &Array2<F>) -> Array2<F> {
    proba.mapv(|p| p.ln())
}