An issue with an associated common check originally discussed in #15015

EDIT: here is a report of statistical worst offenders:

https://github.com/snath-xoc/sample-weight-audit-nondet/blob/main/reports/sklearn_estimators_sample_weight_audit_report.ipynb

This is a pretty simple sample_weight test that says that a weight of 0 is equivalent to not having the samples.
I think every failure here should be considered a bug. This is:

• AdaBoostClassifier/Regressor
• BayesianRidge FIX Sample weight in BayesianRidge #30644
• CalibratedClassifierCV Make check_sample_weights_invariance cv-aware #29796
• CategoricalNB Fix set CategoricalNB().__sklearn_tags__.input_tags.categorical to True #31556
• GradientBoostingClassifier/Regressor
  • not sure if fit is deterministic or not with the config used in check_sample_weight_equivalence: to be investigated.
    • rounding errors make it non-deterministic and break equivalence because of systematic bias to handle near-tied splits: [Meta] consistently break ties randomly in scikit-learn estimators (with random_state) in an unbiased way #23728 (comment)
• HistGradientBoostingClassifier/Regressor
  • check_sample_weight_equivalence fails event when subsampling for binning is not enabled.
  • subsampling for binning needs to be sample_weight aware but can then be only properly tested with a statistical test instead of check_sample_weight_equivalence
  • subsampling also happens (without respecting weights) to compute non-default scoring on a subsampled training set when early stopping is enabled without a validation set
  • Added sample weight handling to BinMapper under HGBT #29641
• HuberRegressor
• LinearRegression
  • Fix LinearRegression's numerical stability on rank deficient data by setting the cond parameter in the call to scipy.linalg.lstsq #30040
  • LinearRegression on sparse matrices is not sample weight consistent #30131
• LinearSVC
  • Fix linear svc handling sample weights under class_weight="balanced" #30057
  • apparently related to the use of liblinear
• LinearSVR
  • apparently related to the use of liblinear
• LogisticRegression
  • lbfgs causes check_sample_weight_equivalence to fail (slightly)
  • liblinear with C=0.01 causes check_sample_weight_equivalence to fail (slightly)
  • Fix sample weight handling in SAG(A) #31675
• LogisticRegressionCV
  • Fix sample weight handling in scoring _log_reg_scoring_path #29419
• ElasticNetCV / LassoCV Fix elasticnect cv sample weight #29442 and Make check_sample_weights_invariance cv-aware #29796 (both are needed)
• Ridge
  • check_sample_weight_equivalence now passes for this estimator after lowering the tol value for lsqr and sparse-cg in the per-check params.
• RidgeClassifier
  • check_sample_weight_equivalence now passes for this estimator after lowering the tol value for lsqr and sparse-cg in the per-check params.
• RidgeClassifierCV. Make check_sample_weights_invariance cv-aware #29796
• DecisionTreeRegressor
  • biased handling of near-tied splits: [Meta] consistently break ties randomly in scikit-learn estimators (with random_state) in an unbiased way #23728 (comment)
• RidgeCV and RidgeClassifierCV can delegate to GridSearchCV when using a non-default cv (which is the case in check_sample_weight_equivalence) or scoring params.
  • fixed by FIX Forward sample weight to the scorer in grid search #30743
• OneClassSVM
• NuSVC (same as SVC)
• NuSVR
• SVC
  • check_sample_weight_equivalence fails with probability=False: to be investigated
  • this expected with probability=True as the weights are not propagated to the internal CV implemented in libsvm
• SVR
• GridSearchCV / RandomizedSearchCV / ...
  • did not forward sample_weight to their scorer by default FIX Forward sample weight to the scorer in grid search #30743.
  • do this even when metadata routing is disabled
  • implement a default routing policy for sample_weight in general: SLEP006: default routing #26179

The following estimators have a stochastic fit, so testing for correct handling of sample weights cannot be tested with check_sample_weight_equivalence but instead requires a statistical test:

• BaggingClassifier/Regressor
  • FIX Draw indices using sample_weight in Bagging #31414
• BisectingKMeans
• KBinsDiscretizer
  • Fix sample weight passing in KBinsDiscretizer #29907
  • note that with a small n_samples and uniform or quantile strategies, the fit is deterministic.
• SGDClassifier/Regressor
• Perceptron (likely same bug as SGD)
• RANSACRegressor
  • RANSACRegressor should use weighted loss functions #15836
• KMeans
• MiniBatchKMeans
  • MiniBatchKMeans not handling sample weights as expected #30750
• IsolationForest
• RandomForestClassifier/Regressor & ExtraTrees
  • FIX Draw indices using sample_weight in Random Forests #31529
• RandomTreesEmbedding

(some might have been fixed since, need to check).

The required sample weight invariance properties (including the behavior of sw=0) were also discussed in #15657

EDIT: expected sample_weight semantics have since been more generally in the refactoring of check_sample_weight_invariance into check_sample_weight_equivalence to check fitting with integer sample weights is equivalent to fitting with repeated data points with a number of repetitions that matches the original weights.