Skip to content

scikit-lego

The scikit-lego pack has 23 rules, and xpd can fix 5 of them.

The scikit-lego rules report arguments that give a wrong number without an error, steps that read the wrong rows or can't be fitted, and code that raises or warns on scikit-lego 0.9.10, in each file that imports sklego. Every number about the scikit-lego rules was measured on scikit-lego 0.9.10 with scikit-learn 1.9.1, NumPy 2.5.5 and pandas 3.1.1. The examples assume that the names they use are imported, and names such as X_train, y_test and model stand for values from your program.

A rule document whose advice needs a newer release of scikit-lego carries since. xpd reads the installed version from the project's venv and skips the document on an older release. When the venv has no scikit-lego, or there is no venv, every document runs.

The pack has no package checks. cvxpy, umap-learn and formulaic are extras that only some estimators need, and a package check can't see which estimators a project uses. When an extra is missing, scikit-lego raises an ImportError that names it.

A rule is an error only when code was run that shows the wrong number or the exception, and its message quotes what was measured. A proof is one of the following kinds:

Proof What runs
number The old and the new form on the same data, and a comparison of what they return.
interpreter The old form, with its exception or its warning under python -W error, and the new form, which runs clean.
speed The old form, timed at two settings.

A rule is a warning when the proof passed and a pattern can't see whether the code wants the old form. A rule is info when no proof of the listed kinds applies.

The rules were then run over scikit-lego's own source, docs scripts and tests at the 0.9.10 tag, 134 Python files, and every report was read. Of the 46 reports there, 42 are on a form that the rule describes, most of them in tests that build a deprecated or superseded class or check that a wrong call raises. The other 4 are on code that is right as it stands:

  • the package's own export of PandasTypeSelector
  • two tests of add_lags with a positive lag
  • one test that passes threshold=0.1 to GMMOutlierDetector only to build an object

After that reading, debug-pipeline-default-log was narrowed to leave a file alone when it imports logging or uses pytest's caplog, which removed 5 reports. unseeded-random-estimator was narrowed to leave test files alone, which removed 7.

Rule Severity Proof Reports on scikit-lego's files On code that is right Fix
thresholder-score error number: 0.887 at three thresholds against 0.740, 0.887 and 0.847 0 0 none
estimator-transformer-proba error number: shape (800, 1) for 400 rows 0 0 none
column-capper-fraction error number: 9,910 of 10,000 values changed against 1,000 0 0 none
dict-mapper-not-int error number: 0.5 and 1.7 became 0 and 1. Interpreter: ValueError 0 0 none
regression-outlier-column-name error number: coefficients 0, 0 and 1, and 11 rows flagged against 10 2 0 none
pca-outlier-after-scaler error number: 320 rows flagged against 20 0 0 none
outlier-threshold error interpreter: ValueError. Number: 380 of 400 rows flagged against 20 1 1 none
grouped-shrinkage-arguments error number: probabilities bit equal with and without the keyword. Interpreter: TypeError 0 0 unsafe
add-lags-positive warning number: accuracy 0.996 against 0.474 on noise 2 2 none
decay-on-shuffled-rows warning number: slope 1.621 against 3.028 0 0 none
outlier-remover-before-model error interpreter: ValueError 0 0 none
decay-without-sample-weight error interpreter: ValueError, TypeError 0 0 none
zero-inflated-roles error interpreter: ValueError 2 0 none
scorer-called-as-metric error interpreter: IndexError, TypeError. Number: a function where a score is expected 0 0 unsafe
dataset-frame-and-xy error interpreter: ValueError 0 0 none
log-step-extra-bare error number: a function where a frame is expected. Interpreter: ValueError 1 0 none
selector-without-feature-names error interpreter: ValueError, AttributeError 0 0 none
debug-pipeline-default-log info number: 0 characters printed 0 0 none
removed-in-sklego error interpreter: ImportError, ModuleNotFoundError, TypeError 0 0 unsafe
deprecated-in-sklego warning interpreter: DeprecationWarning 10 1 safe
superseded-estimator warning, info interpreter: UserWarning 28 0 none
group-time-series-split-folds warning speed: 1.08 s at 5 folds against 10.48 s at 6 0 0 none
unseeded-random-estimator info none of the three 0 0 none

A rule with no report on scikit-lego's files was checked only on the pack's sample files, so its rate of wrong reports on real code is not measured. The pack was not judged on held-back files from other projects, and it has no recall number.

Each rule in the group reports an argument that gives another number than the code asks for, without an error, or that raises where the argument is read.

Rule Severity Fix What it finds
thresholder-score error none Thresholder.score ignores the threshold, so a search over thresholds scores every one the same.
estimator-transformer-proba error none EstimatorTransformer reshapes its output to one column, so predict_proba comes back with twice as many rows and a pipeline around it raises.
column-capper-fraction error none ColumnCapper reads quantile_range in percent, so (0.05, 0.95) caps almost every value at a percentile below 1.
dict-mapper-not-int error none DictMapper casts every mapped value to int, so a float loses its fraction without a warning.
regression-outlier-column-name error none RegressionOutlierDetector with a column name keeps the target among the features, so the inner model predicts the target from itself.
pca-outlier-after-scaler error none After a centring step such as StandardScaler, the default relative error of PCAOutlierDetection divides by numbers near zero and flags many normal rows.
outlier-threshold error none These outlier detectors need a valid threshold, and fit raises ValueError without one or with one out of range.
grouped-shrinkage-arguments error unsafe GroupedClassifier drops a shrinkage keyword without an error, so the setting has no effect.

Each rule in the group reports a step fitted on rows it shouldn't see in that order, or a meta estimator built around a model it can't fit.

Rule Severity Fix What it finds
add-lags-positive warning none add_lags with a positive lag copies the value of a later row, so lag 1 is the next row and leaks the future into the features.
decay-on-shuffled-rows warning none DecayEstimator gives the last rows the most weight, and train_test_split shuffles the rows, so the weights no longer follow time.
outlier-remover-before-model error none OutlierRemover drops rows of X and leaves y whole, so the model after it raises ValueError.
decay-without-sample-weight error none DecayEstimator passes sample_weight to the fit of the model it wraps, so wrapping a Pipeline or a model without sample_weight raises ValueError.
zero-inflated-roles error none ZeroInflatedRegressor takes the classifier first and the regressor second, so swapping them raises ValueError at fit.

Call each scikit-lego scorer, dataset loader and pipeline helper with the arguments it reads, because the forms below raise or return something other than the code expects.

Rule Severity Fix What it finds
scorer-called-as-metric error unsafe These fairness functions return a scorer that takes an estimator, so wrapping them in make_scorer or calling them on labels raises and makes every score NaN.
dataset-frame-and-xy error none With as_frame=True, the scikit-lego loaders ignore return_X_y=True and return one frame with the target in it, so unpacking into X and y raises ValueError.
log-step-extra-bare error none A bare @log_step_extra takes the decorated function as a log function, so calling the function returns another function and not a frame.
selector-without-feature-names error none ColumnSelector, ColumnDropper and TypeSelector have no set_output and no get_feature_names_out, so calling either on a pipeline that holds them raises.
debug-pipeline-default-log info none log_callback="default" logs at level INFO, which Python hides until logging is set up, so the pipeline prints nothing.

Each rule in the group reports a form that raises or warns on scikit-lego 0.9.10, and each removed form raised when it was run. The deprecated and superseded forms were run under python -W error, where the old form raised its warning and the new form ran clean. The messages take each replacement from the text of the warning or from the release that made the change. A fix of a removed form turns code that raises into code that runs, so no such fix is safe.

Rule Severity Fix What it finds
removed-in-sklego error unsafe These forms were removed from scikit-lego, so the import or the call raises.
deprecated-in-sklego warning safe These names, keywords and attributes warn on scikit-lego and will be removed.
superseded-estimator warning unsafe scikit-lego warns on these classes and asks you to use a replacement, such as scikit-learn's QuantileRegressor.

Keep GroupTimeSeriesSplit to few folds, and give each estimator that draws random numbers a seed, so a run finishes and repeats.

Rule Severity Fix What it finds
group-time-series-split-folds warning none GroupTimeSeriesSplit tries every combination of split points, so each extra fold multiplies the time, from about 1 second at 5 folds to 10 at 6.
unseeded-random-estimator info none RandomRegressor and RandomAdder without random_state draw new numbers on every call, so results change between calls.

The pack has no rule for the following habits:

  • MaximumRelevanceMinimumRedundancy with the default kind="auto" scores a whole-number regression target as classes. On one such target it picked the features 0, 1 and 3, against 0, 1 and 4 with kind="regression". A pattern can't see the target, and a rule on every call without kind would report the library's own tests.
  • RegressionOutlierDetector with a column position on a frame, and TimeGapSplit, RandomAdder and OutlierRemover given an input type they don't take, depend on what fit is given, which a pattern can't see.
  • The extras cvxpy, umap-learn and formulaic have no rule, because a package check can't see which estimators a project uses.