File size: 27,693 Bytes
7885a28
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
"""
Various bayesian regression
"""

# Authors: The scikit-learn developers
# SPDX-License-Identifier: BSD-3-Clause

from math import log
from numbers import Integral, Real

import numpy as np
from scipy import linalg
from scipy.linalg import pinvh

from ..base import RegressorMixin, _fit_context
from ..utils import _safe_indexing
from ..utils._param_validation import Interval
from ..utils.extmath import fast_logdet
from ..utils.validation import _check_sample_weight, validate_data
from ._base import LinearModel, _preprocess_data, _rescale_data

###############################################################################
# BayesianRidge regression


class BayesianRidge(RegressorMixin, LinearModel):
    """Bayesian ridge regression.

    Fit a Bayesian ridge model. See the Notes section for details on this
    implementation and the optimization of the regularization parameters
    lambda (precision of the weights) and alpha (precision of the noise).

    Read more in the :ref:`User Guide <bayesian_regression>`.
    For an intuitive visualization of how the sinusoid is approximated by
    a polynomial using different pairs of initial values, see
    :ref:`sphx_glr_auto_examples_linear_model_plot_bayesian_ridge_curvefit.py`.

    Parameters
    ----------
    max_iter : int, default=300
        Maximum number of iterations over the complete dataset before
        stopping independently of any early stopping criterion.

        .. versionchanged:: 1.3

    tol : float, default=1e-3
        Stop the algorithm if w has converged.

    alpha_1 : float, default=1e-6
        Hyper-parameter : shape parameter for the Gamma distribution prior
        over the alpha parameter.

    alpha_2 : float, default=1e-6
        Hyper-parameter : inverse scale parameter (rate parameter) for the
        Gamma distribution prior over the alpha parameter.

    lambda_1 : float, default=1e-6
        Hyper-parameter : shape parameter for the Gamma distribution prior
        over the lambda parameter.

    lambda_2 : float, default=1e-6
        Hyper-parameter : inverse scale parameter (rate parameter) for the
        Gamma distribution prior over the lambda parameter.

    alpha_init : float, default=None
        Initial value for alpha (precision of the noise).
        If not set, alpha_init is 1/Var(y).

        .. versionadded:: 0.22

    lambda_init : float, default=None
        Initial value for lambda (precision of the weights).
        If not set, lambda_init is 1.

        .. versionadded:: 0.22

    compute_score : bool, default=False
        If True, compute the log marginal likelihood at each iteration of the
        optimization.

    fit_intercept : bool, default=True
        Whether to calculate the intercept for this model.
        The intercept is not treated as a probabilistic parameter
        and thus has no associated variance. If set
        to False, no intercept will be used in calculations
        (i.e. data is expected to be centered).

    copy_X : bool, default=True
        If True, X will be copied; else, it may be overwritten.

    verbose : bool, default=False
        Verbose mode when fitting the model.

    Attributes
    ----------
    coef_ : array-like of shape (n_features,)
        Coefficients of the regression model (mean of distribution)

    intercept_ : float
        Independent term in decision function. Set to 0.0 if
        `fit_intercept = False`.

    alpha_ : float
       Estimated precision of the noise.

    lambda_ : float
       Estimated precision of the weights.

    sigma_ : array-like of shape (n_features, n_features)
        Estimated variance-covariance matrix of the weights

    scores_ : array-like of shape (n_iter_+1,)
        If computed_score is True, value of the log marginal likelihood (to be
        maximized) at each iteration of the optimization. The array starts
        with the value of the log marginal likelihood obtained for the initial
        values of alpha and lambda and ends with the value obtained for the
        estimated alpha and lambda.

    n_iter_ : int
        The actual number of iterations to reach the stopping criterion.

    X_offset_ : ndarray of shape (n_features,)
        If `fit_intercept=True`, offset subtracted for centering data to a
        zero mean. Set to np.zeros(n_features) otherwise.

    X_scale_ : ndarray of shape (n_features,)
        Set to np.ones(n_features).

    n_features_in_ : int
        Number of features seen during :term:`fit`.

        .. versionadded:: 0.24

    feature_names_in_ : ndarray of shape (`n_features_in_`,)
        Names of features seen during :term:`fit`. Defined only when `X`
        has feature names that are all strings.

        .. versionadded:: 1.0

    See Also
    --------
    ARDRegression : Bayesian ARD regression.

    Notes
    -----
    There exist several strategies to perform Bayesian ridge regression. This
    implementation is based on the algorithm described in Appendix A of
    (Tipping, 2001) where updates of the regularization parameters are done as
    suggested in (MacKay, 1992). Note that according to A New
    View of Automatic Relevance Determination (Wipf and Nagarajan, 2008) these
    update rules do not guarantee that the marginal likelihood is increasing
    between two consecutive iterations of the optimization.

    References
    ----------
    D. J. C. MacKay, Bayesian Interpolation, Computation and Neural Systems,
    Vol. 4, No. 3, 1992.

    M. E. Tipping, Sparse Bayesian Learning and the Relevance Vector Machine,
    Journal of Machine Learning Research, Vol. 1, 2001.

    Examples
    --------
    >>> from sklearn import linear_model
    >>> clf = linear_model.BayesianRidge()
    >>> clf.fit([[0,0], [1, 1], [2, 2]], [0, 1, 2])
    BayesianRidge()
    >>> clf.predict([[1, 1]])
    array([1.])
    """

    _parameter_constraints: dict = {
        "max_iter": [Interval(Integral, 1, None, closed="left")],
        "tol": [Interval(Real, 0, None, closed="neither")],
        "alpha_1": [Interval(Real, 0, None, closed="left")],
        "alpha_2": [Interval(Real, 0, None, closed="left")],
        "lambda_1": [Interval(Real, 0, None, closed="left")],
        "lambda_2": [Interval(Real, 0, None, closed="left")],
        "alpha_init": [None, Interval(Real, 0, None, closed="left")],
        "lambda_init": [None, Interval(Real, 0, None, closed="left")],
        "compute_score": ["boolean"],
        "fit_intercept": ["boolean"],
        "copy_X": ["boolean"],
        "verbose": ["verbose"],
    }

    def __init__(
        self,
        *,
        max_iter=300,
        tol=1.0e-3,
        alpha_1=1.0e-6,
        alpha_2=1.0e-6,
        lambda_1=1.0e-6,
        lambda_2=1.0e-6,
        alpha_init=None,
        lambda_init=None,
        compute_score=False,
        fit_intercept=True,
        copy_X=True,
        verbose=False,
    ):
        self.max_iter = max_iter
        self.tol = tol
        self.alpha_1 = alpha_1
        self.alpha_2 = alpha_2
        self.lambda_1 = lambda_1
        self.lambda_2 = lambda_2
        self.alpha_init = alpha_init
        self.lambda_init = lambda_init
        self.compute_score = compute_score
        self.fit_intercept = fit_intercept
        self.copy_X = copy_X
        self.verbose = verbose

    @_fit_context(prefer_skip_nested_validation=True)
    def fit(self, X, y, sample_weight=None):
        """Fit the model.

        Parameters
        ----------
        X : ndarray of shape (n_samples, n_features)
            Training data.
        y : ndarray of shape (n_samples,)
            Target values. Will be cast to X's dtype if necessary.

        sample_weight : ndarray of shape (n_samples,), default=None
            Individual weights for each sample.

            .. versionadded:: 0.20
               parameter *sample_weight* support to BayesianRidge.

        Returns
        -------
        self : object
            Returns the instance itself.
        """
        X, y = validate_data(
            self,
            X,
            y,
            dtype=[np.float64, np.float32],
            force_writeable=True,
            y_numeric=True,
        )
        dtype = X.dtype

        if sample_weight is not None:
            sample_weight = _check_sample_weight(sample_weight, X, dtype=dtype)

        X, y, X_offset_, y_offset_, X_scale_ = _preprocess_data(
            X,
            y,
            fit_intercept=self.fit_intercept,
            copy=self.copy_X,
            sample_weight=sample_weight,
        )

        if sample_weight is not None:
            # Sample weight can be implemented via a simple rescaling.
            X, y, _ = _rescale_data(X, y, sample_weight)

        self.X_offset_ = X_offset_
        self.X_scale_ = X_scale_
        n_samples, n_features = X.shape

        # Initialization of the values of the parameters
        eps = np.finfo(np.float64).eps
        # Add `eps` in the denominator to omit division by zero if `np.var(y)`
        # is zero
        alpha_ = self.alpha_init
        lambda_ = self.lambda_init
        if alpha_ is None:
            alpha_ = 1.0 / (np.var(y) + eps)
        if lambda_ is None:
            lambda_ = 1.0

        # Avoid unintended type promotion to float64 with numpy 2
        alpha_ = np.asarray(alpha_, dtype=dtype)
        lambda_ = np.asarray(lambda_, dtype=dtype)

        verbose = self.verbose
        lambda_1 = self.lambda_1
        lambda_2 = self.lambda_2
        alpha_1 = self.alpha_1
        alpha_2 = self.alpha_2

        self.scores_ = list()
        coef_old_ = None

        XT_y = np.dot(X.T, y)
        U, S, Vh = linalg.svd(X, full_matrices=False)
        eigen_vals_ = S**2

        # Convergence loop of the bayesian ridge regression
        for iter_ in range(self.max_iter):
            # update posterior mean coef_ based on alpha_ and lambda_ and
            # compute corresponding rmse
            coef_, rmse_ = self._update_coef_(
                X, y, n_samples, n_features, XT_y, U, Vh, eigen_vals_, alpha_, lambda_
            )
            if self.compute_score:
                # compute the log marginal likelihood
                s = self._log_marginal_likelihood(
                    n_samples, n_features, eigen_vals_, alpha_, lambda_, coef_, rmse_
                )
                self.scores_.append(s)

            # Update alpha and lambda according to (MacKay, 1992)
            gamma_ = np.sum((alpha_ * eigen_vals_) / (lambda_ + alpha_ * eigen_vals_))
            lambda_ = (gamma_ + 2 * lambda_1) / (np.sum(coef_**2) + 2 * lambda_2)
            alpha_ = (n_samples - gamma_ + 2 * alpha_1) / (rmse_ + 2 * alpha_2)

            # Check for convergence
            if iter_ != 0 and np.sum(np.abs(coef_old_ - coef_)) < self.tol:
                if verbose:
                    print("Convergence after ", str(iter_), " iterations")
                break
            coef_old_ = np.copy(coef_)

        self.n_iter_ = iter_ + 1

        # return regularization parameters and corresponding posterior mean,
        # log marginal likelihood and posterior covariance
        self.alpha_ = alpha_
        self.lambda_ = lambda_
        self.coef_, rmse_ = self._update_coef_(
            X, y, n_samples, n_features, XT_y, U, Vh, eigen_vals_, alpha_, lambda_
        )
        if self.compute_score:
            # compute the log marginal likelihood
            s = self._log_marginal_likelihood(
                n_samples, n_features, eigen_vals_, alpha_, lambda_, coef_, rmse_
            )
            self.scores_.append(s)
            self.scores_ = np.array(self.scores_)

        # posterior covariance is given by 1/alpha_ * scaled_sigma_
        scaled_sigma_ = np.dot(
            Vh.T, Vh / (eigen_vals_ + lambda_ / alpha_)[:, np.newaxis]
        )
        self.sigma_ = (1.0 / alpha_) * scaled_sigma_

        self._set_intercept(X_offset_, y_offset_, X_scale_)

        return self

    def predict(self, X, return_std=False):
        """Predict using the linear model.

        In addition to the mean of the predictive distribution, also its
        standard deviation can be returned.

        Parameters
        ----------
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Samples.

        return_std : bool, default=False
            Whether to return the standard deviation of posterior prediction.

        Returns
        -------
        y_mean : array-like of shape (n_samples,)
            Mean of predictive distribution of query points.

        y_std : array-like of shape (n_samples,)
            Standard deviation of predictive distribution of query points.
        """
        y_mean = self._decision_function(X)
        if not return_std:
            return y_mean
        else:
            sigmas_squared_data = (np.dot(X, self.sigma_) * X).sum(axis=1)
            y_std = np.sqrt(sigmas_squared_data + (1.0 / self.alpha_))
            return y_mean, y_std

    def _update_coef_(
        self, X, y, n_samples, n_features, XT_y, U, Vh, eigen_vals_, alpha_, lambda_
    ):
        """Update posterior mean and compute corresponding rmse.

        Posterior mean is given by coef_ = scaled_sigma_ * X.T * y where
        scaled_sigma_ = (lambda_/alpha_ * np.eye(n_features)
                         + np.dot(X.T, X))^-1
        """

        if n_samples > n_features:
            coef_ = np.linalg.multi_dot(
                [Vh.T, Vh / (eigen_vals_ + lambda_ / alpha_)[:, np.newaxis], XT_y]
            )
        else:
            coef_ = np.linalg.multi_dot(
                [X.T, U / (eigen_vals_ + lambda_ / alpha_)[None, :], U.T, y]
            )

        rmse_ = np.sum((y - np.dot(X, coef_)) ** 2)

        return coef_, rmse_

    def _log_marginal_likelihood(
        self, n_samples, n_features, eigen_vals, alpha_, lambda_, coef, rmse
    ):
        """Log marginal likelihood."""
        alpha_1 = self.alpha_1
        alpha_2 = self.alpha_2
        lambda_1 = self.lambda_1
        lambda_2 = self.lambda_2

        # compute the log of the determinant of the posterior covariance.
        # posterior covariance is given by
        # sigma = (lambda_ * np.eye(n_features) + alpha_ * np.dot(X.T, X))^-1
        if n_samples > n_features:
            logdet_sigma = -np.sum(np.log(lambda_ + alpha_ * eigen_vals))
        else:
            logdet_sigma = np.full(n_features, lambda_, dtype=np.array(lambda_).dtype)
            logdet_sigma[:n_samples] += alpha_ * eigen_vals
            logdet_sigma = -np.sum(np.log(logdet_sigma))

        score = lambda_1 * log(lambda_) - lambda_2 * lambda_
        score += alpha_1 * log(alpha_) - alpha_2 * alpha_
        score += 0.5 * (
            n_features * log(lambda_)
            + n_samples * log(alpha_)
            - alpha_ * rmse
            - lambda_ * np.sum(coef**2)
            + logdet_sigma
            - n_samples * log(2 * np.pi)
        )

        return score


###############################################################################
# ARD (Automatic Relevance Determination) regression


class ARDRegression(RegressorMixin, LinearModel):
    """Bayesian ARD regression.

    Fit the weights of a regression model, using an ARD prior. The weights of
    the regression model are assumed to be in Gaussian distributions.
    Also estimate the parameters lambda (precisions of the distributions of the
    weights) and alpha (precision of the distribution of the noise).
    The estimation is done by an iterative procedures (Evidence Maximization)

    Read more in the :ref:`User Guide <bayesian_regression>`.

    Parameters
    ----------
    max_iter : int, default=300
        Maximum number of iterations.

        .. versionchanged:: 1.3

    tol : float, default=1e-3
        Stop the algorithm if w has converged.

    alpha_1 : float, default=1e-6
        Hyper-parameter : shape parameter for the Gamma distribution prior
        over the alpha parameter.

    alpha_2 : float, default=1e-6
        Hyper-parameter : inverse scale parameter (rate parameter) for the
        Gamma distribution prior over the alpha parameter.

    lambda_1 : float, default=1e-6
        Hyper-parameter : shape parameter for the Gamma distribution prior
        over the lambda parameter.

    lambda_2 : float, default=1e-6
        Hyper-parameter : inverse scale parameter (rate parameter) for the
        Gamma distribution prior over the lambda parameter.

    compute_score : bool, default=False
        If True, compute the objective function at each step of the model.

    threshold_lambda : float, default=10 000
        Threshold for removing (pruning) weights with high precision from
        the computation.

    fit_intercept : bool, default=True
        Whether to calculate the intercept for this model. If set
        to false, no intercept will be used in calculations
        (i.e. data is expected to be centered).

    copy_X : bool, default=True
        If True, X will be copied; else, it may be overwritten.

    verbose : bool, default=False
        Verbose mode when fitting the model.

    Attributes
    ----------
    coef_ : array-like of shape (n_features,)
        Coefficients of the regression model (mean of distribution)

    alpha_ : float
       estimated precision of the noise.

    lambda_ : array-like of shape (n_features,)
       estimated precisions of the weights.

    sigma_ : array-like of shape (n_features, n_features)
        estimated variance-covariance matrix of the weights

    scores_ : float
        if computed, value of the objective function (to be maximized)

    n_iter_ : int
        The actual number of iterations to reach the stopping criterion.

        .. versionadded:: 1.3

    intercept_ : float
        Independent term in decision function. Set to 0.0 if
        ``fit_intercept = False``.

    X_offset_ : float
        If `fit_intercept=True`, offset subtracted for centering data to a
        zero mean. Set to np.zeros(n_features) otherwise.

    X_scale_ : float
        Set to np.ones(n_features).

    n_features_in_ : int
        Number of features seen during :term:`fit`.

        .. versionadded:: 0.24

    feature_names_in_ : ndarray of shape (`n_features_in_`,)
        Names of features seen during :term:`fit`. Defined only when `X`
        has feature names that are all strings.

        .. versionadded:: 1.0

    See Also
    --------
    BayesianRidge : Bayesian ridge regression.

    Notes
    -----
    For an example, see :ref:`examples/linear_model/plot_ard.py
    <sphx_glr_auto_examples_linear_model_plot_ard.py>`.

    References
    ----------
    D. J. C. MacKay, Bayesian nonlinear modeling for the prediction
    competition, ASHRAE Transactions, 1994.

    R. Salakhutdinov, Lecture notes on Statistical Machine Learning,
    http://www.utstat.toronto.edu/~rsalakhu/sta4273/notes/Lecture2.pdf#page=15
    Their beta is our ``self.alpha_``
    Their alpha is our ``self.lambda_``
    ARD is a little different than the slide: only dimensions/features for
    which ``self.lambda_ < self.threshold_lambda`` are kept and the rest are
    discarded.

    Examples
    --------
    >>> from sklearn import linear_model
    >>> clf = linear_model.ARDRegression()
    >>> clf.fit([[0,0], [1, 1], [2, 2]], [0, 1, 2])
    ARDRegression()
    >>> clf.predict([[1, 1]])
    array([1.])
    """

    _parameter_constraints: dict = {
        "max_iter": [Interval(Integral, 1, None, closed="left")],
        "tol": [Interval(Real, 0, None, closed="left")],
        "alpha_1": [Interval(Real, 0, None, closed="left")],
        "alpha_2": [Interval(Real, 0, None, closed="left")],
        "lambda_1": [Interval(Real, 0, None, closed="left")],
        "lambda_2": [Interval(Real, 0, None, closed="left")],
        "compute_score": ["boolean"],
        "threshold_lambda": [Interval(Real, 0, None, closed="left")],
        "fit_intercept": ["boolean"],
        "copy_X": ["boolean"],
        "verbose": ["verbose"],
    }

    def __init__(
        self,
        *,
        max_iter=300,
        tol=1.0e-3,
        alpha_1=1.0e-6,
        alpha_2=1.0e-6,
        lambda_1=1.0e-6,
        lambda_2=1.0e-6,
        compute_score=False,
        threshold_lambda=1.0e4,
        fit_intercept=True,
        copy_X=True,
        verbose=False,
    ):
        self.max_iter = max_iter
        self.tol = tol
        self.fit_intercept = fit_intercept
        self.alpha_1 = alpha_1
        self.alpha_2 = alpha_2
        self.lambda_1 = lambda_1
        self.lambda_2 = lambda_2
        self.compute_score = compute_score
        self.threshold_lambda = threshold_lambda
        self.copy_X = copy_X
        self.verbose = verbose

    @_fit_context(prefer_skip_nested_validation=True)
    def fit(self, X, y):
        """Fit the model according to the given training data and parameters.

        Iterative procedure to maximize the evidence

        Parameters
        ----------
        X : array-like of shape (n_samples, n_features)
            Training vector, where `n_samples` is the number of samples and
            `n_features` is the number of features.
        y : array-like of shape (n_samples,)
            Target values (integers). Will be cast to X's dtype if necessary.

        Returns
        -------
        self : object
            Fitted estimator.
        """
        X, y = validate_data(
            self,
            X,
            y,
            dtype=[np.float64, np.float32],
            force_writeable=True,
            y_numeric=True,
            ensure_min_samples=2,
        )
        dtype = X.dtype

        n_samples, n_features = X.shape
        coef_ = np.zeros(n_features, dtype=dtype)

        X, y, X_offset_, y_offset_, X_scale_ = _preprocess_data(
            X, y, fit_intercept=self.fit_intercept, copy=self.copy_X
        )

        self.X_offset_ = X_offset_
        self.X_scale_ = X_scale_

        # Launch the convergence loop
        keep_lambda = np.ones(n_features, dtype=bool)

        lambda_1 = self.lambda_1
        lambda_2 = self.lambda_2
        alpha_1 = self.alpha_1
        alpha_2 = self.alpha_2
        verbose = self.verbose

        # Initialization of the values of the parameters
        eps = np.finfo(np.float64).eps
        # Add `eps` in the denominator to omit division by zero if `np.var(y)`
        # is zero.
        # Explicitly set dtype to avoid unintended type promotion with numpy 2.
        alpha_ = np.asarray(1.0 / (np.var(y) + eps), dtype=dtype)
        lambda_ = np.ones(n_features, dtype=dtype)

        self.scores_ = list()
        coef_old_ = None

        def update_coeff(X, y, coef_, alpha_, keep_lambda, sigma_):
            coef_[keep_lambda] = alpha_ * np.linalg.multi_dot(
                [sigma_, X[:, keep_lambda].T, y]
            )
            return coef_

        update_sigma = (
            self._update_sigma
            if n_samples >= n_features
            else self._update_sigma_woodbury
        )
        # Iterative procedure of ARDRegression
        for iter_ in range(self.max_iter):
            sigma_ = update_sigma(X, alpha_, lambda_, keep_lambda)
            coef_ = update_coeff(X, y, coef_, alpha_, keep_lambda, sigma_)

            # Update alpha and lambda
            rmse_ = np.sum((y - np.dot(X, coef_)) ** 2)
            gamma_ = 1.0 - lambda_[keep_lambda] * np.diag(sigma_)
            lambda_[keep_lambda] = (gamma_ + 2.0 * lambda_1) / (
                (coef_[keep_lambda]) ** 2 + 2.0 * lambda_2
            )
            alpha_ = (n_samples - gamma_.sum() + 2.0 * alpha_1) / (
                rmse_ + 2.0 * alpha_2
            )

            # Prune the weights with a precision over a threshold
            keep_lambda = lambda_ < self.threshold_lambda
            coef_[~keep_lambda] = 0

            # Compute the objective function
            if self.compute_score:
                s = (lambda_1 * np.log(lambda_) - lambda_2 * lambda_).sum()
                s += alpha_1 * log(alpha_) - alpha_2 * alpha_
                s += 0.5 * (
                    fast_logdet(sigma_)
                    + n_samples * log(alpha_)
                    + np.sum(np.log(lambda_))
                )
                s -= 0.5 * (alpha_ * rmse_ + (lambda_ * coef_**2).sum())
                self.scores_.append(s)

            # Check for convergence
            if iter_ > 0 and np.sum(np.abs(coef_old_ - coef_)) < self.tol:
                if verbose:
                    print("Converged after %s iterations" % iter_)
                break
            coef_old_ = np.copy(coef_)

            if not keep_lambda.any():
                break

        self.n_iter_ = iter_ + 1

        if keep_lambda.any():
            # update sigma and mu using updated params from the last iteration
            sigma_ = update_sigma(X, alpha_, lambda_, keep_lambda)
            coef_ = update_coeff(X, y, coef_, alpha_, keep_lambda, sigma_)
        else:
            sigma_ = np.array([]).reshape(0, 0)

        self.coef_ = coef_
        self.alpha_ = alpha_
        self.sigma_ = sigma_
        self.lambda_ = lambda_
        self._set_intercept(X_offset_, y_offset_, X_scale_)
        return self

    def _update_sigma_woodbury(self, X, alpha_, lambda_, keep_lambda):
        # See slides as referenced in the docstring note
        # this function is used when n_samples < n_features and will invert
        # a matrix of shape (n_samples, n_samples) making use of the
        # woodbury formula:
        # https://en.wikipedia.org/wiki/Woodbury_matrix_identity
        n_samples = X.shape[0]
        X_keep = X[:, keep_lambda]
        inv_lambda = 1 / lambda_[keep_lambda].reshape(1, -1)
        sigma_ = pinvh(
            np.eye(n_samples, dtype=X.dtype) / alpha_
            + np.dot(X_keep * inv_lambda, X_keep.T)
        )
        sigma_ = np.dot(sigma_, X_keep * inv_lambda)
        sigma_ = -np.dot(inv_lambda.reshape(-1, 1) * X_keep.T, sigma_)
        sigma_[np.diag_indices(sigma_.shape[1])] += 1.0 / lambda_[keep_lambda]
        return sigma_

    def _update_sigma(self, X, alpha_, lambda_, keep_lambda):
        # See slides as referenced in the docstring note
        # this function is used when n_samples >= n_features and will
        # invert a matrix of shape (n_features, n_features)
        X_keep = X[:, keep_lambda]
        gram = np.dot(X_keep.T, X_keep)
        eye = np.eye(gram.shape[0], dtype=X.dtype)
        sigma_inv = lambda_[keep_lambda] * eye + alpha_ * gram
        sigma_ = pinvh(sigma_inv)
        return sigma_

    def predict(self, X, return_std=False):
        """Predict using the linear model.

        In addition to the mean of the predictive distribution, also its
        standard deviation can be returned.

        Parameters
        ----------
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Samples.

        return_std : bool, default=False
            Whether to return the standard deviation of posterior prediction.

        Returns
        -------
        y_mean : array-like of shape (n_samples,)
            Mean of predictive distribution of query points.

        y_std : array-like of shape (n_samples,)
            Standard deviation of predictive distribution of query points.
        """
        y_mean = self._decision_function(X)
        if return_std is False:
            return y_mean
        else:
            col_index = self.lambda_ < self.threshold_lambda
            X = _safe_indexing(X, indices=col_index, axis=1)
            sigmas_squared_data = (np.dot(X, self.sigma_) * X).sum(axis=1)
            y_std = np.sqrt(sigmas_squared_data + (1.0 / self.alpha_))
            return y_mean, y_std