#!/usr/bin/env python3
"""Lint scikit-learn code for API misuse that silently produces wrong numbers or breaks in production.
(For data leakage use target-leakage-detector/scripts/code_leak_lint.py.)
Rules:
SK001 high ranking/probability metric fed hard labels: roc_auc_score / average_precision_score /
log_loss / brier_score_loss(..., model.predict(X)) -> use predict_proba(X)[:, 1]
SK002 high removed or renamed parameters that raise in current scikit-learn: OneHotEncoder(sparse=),
LogisticRegression(penalty="none"), normalize= in linear models, base_estimator=,
max_features="auto", criterion="mse"/"mae", loss="deviance"/"ls"/"log", affinity=,
mean_squared_error(squared=False), fit_params= in cross_val_score, algorithm="SAMME.R"
SK003 medium LabelEncoder used on feature columns (it is for targets; use OrdinalEncoder/OneHotEncoder)
SK004 medium pd.get_dummies in training code: columns differ between train and inference
SK005 medium FunctionTransformer(lambda ...): lambdas cannot be pickled with the model
SK006 medium model scored on the same data it was fitted on (.fit(X, y) then .score(X, y) / metric on
predict(X)) - reports training performance
SK007 low stochastic estimator / splitter / search without random_state
SK008 low n_jobs=-1 on both a search and its estimator (nested parallelism oversubscribes cores)
SK009 low LogisticRegression without max_iter on unscaled data risks ConvergenceWarning
SK010 info ColumnTransformer without remainder=: unlisted columns are dropped silently
Usage:
python pipeline_lint.py train.py
python pipeline_lint.py src/*.py --json
Exit codes: 0 no high findings, 1 high findings, 2 read or syntax error.
"""
import argparse
import ast
import json
import sys
PROB_METRICS = {"roc_auc_score", "average_precision_score", "log_loss", "brier_score_loss"}
LABEL_METRICS = {"accuracy_score", "f1_score", "precision_score", "recall_score", "r2_score",
"mean_absolute_error", "mean_squared_error", "balanced_accuracy_score"}
STOCHASTIC = {"RandomForestClassifier", "RandomForestRegressor", "ExtraTreesClassifier", "ExtraTreesRegressor",
"GradientBoostingClassifier", "GradientBoostingRegressor", "HistGradientBoostingClassifier",
"HistGradientBoostingRegressor", "KMeans", "MiniBatchKMeans", "IsolationForest", "SGDClassifier",
"SGDRegressor", "MLPClassifier", "MLPRegressor", "train_test_split", "RandomizedSearchCV",
"HalvingRandomSearchCV", "TSNE", "BaggingClassifier", "BaggingRegressor", "XGBClassifier",
"XGBRegressor", "LGBMClassifier", "LGBMRegressor", "CatBoostClassifier", "CatBoostRegressor"}
STOCH_KW = {"LGBMClassifier": "random_state", "LGBMRegressor": "random_state", "CatBoostClassifier": "random_seed",
"CatBoostRegressor": "random_seed"}
SEARCHES = {"GridSearchCV", "RandomizedSearchCV", "HalvingGridSearchCV", "HalvingRandomSearchCV"}
LINEAR = {"LinearRegression", "Ridge", "Lasso", "ElasticNet", "Lars", "LassoLars", "OrthogonalMatchingPursuit"}
def name(node):
f = node.func if isinstance(node, ast.Call) else node
if isinstance(f, ast.Name):
return f.id
if isinstance(f, ast.Attribute):
return f.attr
return ""
def src(node):
try:
return ast.unparse(node)
except Exception:
return ""
def kw(call, key):
for k in call.keywords:
if k.arg == key:
return k.value
return None
def const(v):
if isinstance(v, ast.UnaryOp) and isinstance(v.op, ast.USub) and isinstance(v.operand, ast.Constant):
return -v.operand.value # -1 parses as USub(1)
return v.value if isinstance(v, ast.Constant) else None
def lint(code, path):
tree = ast.parse(code, filename=path)
out = []
def add(rule, sev, node, msg):
out.append({"file": path, "line": node.lineno, "code": rule, "severity": sev, "message": msg})
calls = sorted((n for n in ast.walk(tree) if isinstance(n, ast.Call)), key=lambda n: (n.lineno, n.col_offset))
fitted = {} # data arg text -> line of .fit
for c in calls:
n = name(c)
# SK001
if n in PROB_METRICS and len(c.args) >= 2:
a2 = c.args[1]
if isinstance(a2, ast.Call) and name(a2) == "predict":
add("SK001", "high", c, f"{n}() given hard labels from .predict(); use .predict_proba(X)[:, 1] "
"(or decision_function) - the number is wrong, not just worse")
# SK002
if n == "OneHotEncoder" and kw(c, "sparse") is not None:
add("SK002", "high", c, "OneHotEncoder(sparse=) was removed in 1.4; use sparse_output=")
if n == "LogisticRegression" and const(kw(c, "penalty")) == "none":
add("SK002", "high", c, 'penalty="none" was removed; use penalty=None')
if n in LINEAR and kw(c, "normalize") is not None:
add("SK002", "high", c, "normalize= was removed (1.2); put StandardScaler in a Pipeline")
if kw(c, "base_estimator") is not None:
add("SK002", "high", c, "base_estimator= was renamed to estimator= (removed in 1.4)")
if const(kw(c, "max_features")) == "auto":
add("SK002", "high", c, 'max_features="auto" was removed (1.3); use "sqrt" (classifier) or 1.0 (regressor)')
if const(kw(c, "criterion")) in ("mse", "mae"):
add("SK002", "high", c, 'criterion="mse"/"mae" were removed; use "squared_error"/"absolute_error"')
if const(kw(c, "loss")) in ("deviance", "ls", "lad") or (n.startswith("SGD") and const(kw(c, "loss")) == "log"):
add("SK002", "high", c, f'loss="{const(kw(c, "loss"))}" was removed; use "log_loss", "squared_error" '
'or "absolute_error"')
if kw(c, "affinity") is not None and n in ("AgglomerativeClustering", "FeatureAgglomeration"):
add("SK002", "high", c, "affinity= was removed (1.4); use metric=")
if n == "mean_squared_error" and kw(c, "squared") is not None:
add("SK002", "high", c, "squared= was removed (1.6); use root_mean_squared_error()")
if n in ("cross_val_score", "cross_validate", "cross_val_predict") and kw(c, "fit_params") is not None:
add("SK002", "high", c, "fit_params= was removed (1.6); use params=")
if const(kw(c, "algorithm")) == "SAMME.R":
add("SK002", "high", c, 'AdaBoost algorithm="SAMME.R" is gone (deprecated 1.4; the algorithm parameter itself was later removed): drop it')
# SK003
if n in ("fit_transform", "fit") and isinstance(c.func, ast.Attribute):
recv = c.func.value
if (isinstance(recv, ast.Call) and name(recv) == "LabelEncoder") and c.args and \
not src(c.args[0]).lower().startswith(("y", "target", "label")):
add("SK003", "medium", c, "LabelEncoder on a feature: it is meant for targets and cannot handle unseen "
"values; use OrdinalEncoder(handle_unknown=...) or OneHotEncoder inside the Pipeline")
# SK004
if n == "get_dummies":
add("SK004", "medium", c, "pd.get_dummies: train and inference columns diverge when levels differ; use "
"OneHotEncoder(handle_unknown='ignore') in the Pipeline")
# SK005
if n == "FunctionTransformer" and c.args and isinstance(c.args[0], ast.Lambda) or (
n == "FunctionTransformer" and isinstance(kw(c, "func"), ast.Lambda)):
add("SK005", "medium", c, "FunctionTransformer(lambda ...) cannot be pickled with joblib/pickle; use a "
"named module-level function")
# SK006
if n == "fit" and len(c.args) >= 1:
fitted.setdefault(src(c.args[0]), c.lineno)
if n == "score" and c.args and src(c.args[0]) in fitted and "train" not in src(c.args[0]).lower():
add("SK006", "medium", c, f".score({src(c.args[0])}, ...) on the data the model was fitted on: this is "
"training performance; score a held-out split or use cross-validation")
if n in (PROB_METRICS | LABEL_METRICS) and len(c.args) >= 2:
pred = c.args[1]
if isinstance(pred, ast.Call) and name(pred) in ("predict", "predict_proba") and pred.args and \
src(pred.args[0]) in fitted and "train" not in src(pred.args[0]).lower():
add("SK006", "medium", c, f"{n}() computed on predictions for {src(pred.args[0])}, the data the model "
"was fitted on")
# SK007
if n in STOCHASTIC:
key = STOCH_KW.get(n, "random_state")
if kw(c, key) is None and not (n.startswith("XGB") and kw(c, "seed") is not None):
add("SK007", "low", c, f"{n} without {key}=: results change on every run")
if n in ("KFold", "StratifiedKFold") and const(kw(c, "shuffle")) is True and kw(c, "random_state") is None:
add("SK007", "low", c, f"{n}(shuffle=True) without random_state: folds change on every run")
# SK008
if n in SEARCHES and const(kw(c, "n_jobs")) == -1 and c.args:
est = c.args[0]
if isinstance(est, ast.Call) and const(kw(est, "n_jobs")) == -1:
add("SK008", "low", c, "n_jobs=-1 on the search and its estimator: nested parallelism oversubscribes; "
"parallelise one level")
# SK009
if n == "LogisticRegression" and kw(c, "max_iter") is None:
add("SK009", "low", c, "LogisticRegression without max_iter (default 100): scale features and/or raise "
"max_iter to avoid ConvergenceWarning and an under-fitted model")
# SK010
if n == "ColumnTransformer" and kw(c, "remainder") is None:
add("SK010", "info", c, "ColumnTransformer drops unlisted columns by default; set remainder= explicitly")
uniq = {(f["line"], f["code"]): f for f in out}
return sorted(uniq.values(), key=lambda f: (f["line"], f["code"]))
def main():
ap = argparse.ArgumentParser(description="Lint scikit-learn code for API misuse.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("files", nargs="+")
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
findings = []
for p in a.files:
try:
code = sys.stdin.read() if p == "-" else open(p, encoding="utf-8").read()
findings += lint(code, "<stdin>" if p == "-" else p)
except (OSError, SyntaxError) as e:
print(f"error: {p}: {e}", file=sys.stderr)
return 2
high = sum(f["severity"] == "high" for f in findings)
if a.json:
print(json.dumps({"findings": findings, "high": high}, indent=2))
else:
for f in findings:
print(f"{f['file']}:{f['line']}: {f['code']} [{f['severity']}] {f['message']}")
print(f"\n{len(findings)} finding(s), {high} high" + ("" if findings else ": no scikit-learn API issues found"))
return 1 if high else 0
if __name__ == "__main__":
sys.exit(main())
anomaly-detection-model-builder
Build and evaluate anomaly detection for tabular, sensor and transaction data with few or no labels: define the anomaly type (point, contextual, collective), choose detectors (robust z-score/MAD and IQR rules, Mahalanobis with robust covariance, k-NN distance, LOF, Isolation Forest, one-class SVM, PCA reconstruction / autoencoders, forecasting residuals for time series), combine them by rank, explain each alert by its deviating features, set thresholds from an alert budget or the score distribution, and evaluate with precision@k, PR-AUC and recall on known incidents, or detector agreement when no labels exist. Includes pure-Python detectors (robust z, Mahalanobis, kNN, Isolation Forest, ensemble) and an evaluation script.
catboost-model-builder
Build CatBoost models that use its strengths correctly: raw categorical columns as cat_features (ordered target statistics, no manual encoding), text_features, Pool construction with NaN-safe categories, ordered boosting and has_time for temporal data, symmetric (oblivious) trees, overfitting detector with eval_set and use_best_model, losses for classification, regression (Quantile, Poisson, Tweedie) and ranking (YetiRank), class weights, monotone constraints, feature importance (PredictionValuesChange, LossFunctionChange) and SHAP values, GPU training and model export. Plans cat/text/numeric columns from a CSV and generates a runnable training script, and checks parameter sets for the conflicts CatBoost rejects (iterations vs n_estimators, Bayesian bootstrap with subsample, multiple class-weight options, depth > 16, use_best_model without eval_set).
categorical-encoding-selector
Choose and implement the right encoding for every categorical column given the model family and the data: one-hot (with infrequent-level grouping and handle_unknown), ordinal with an explicit order, cross-fitted target encoding (scikit-learn TargetEncoder, CatBoost ordered statistics, category_encoders inside CV), frequency/count encoding, feature hashing, native categorical support in LightGBM, HistGradientBoosting, XGBoost and CatBoost, prefix levels for codes (zip3, NAICS-2), and embeddings. Measures cardinality, rare-level share, unseen-level rate on a time-ordered holdout and target signal per column, emits the ColumnTransformer, and quantifies target-encoding leakage (in-sample vs out-of-fold vs holdout AUC).
classification-model-selector
Choose the right classification algorithm for tabular, text or mixed data and prove the choice with a fair bake-off: logistic regression (L2, L1, elastic net), linear SVM, naive Bayes, k-NN, RBF SVM, decision trees, random forest, extra trees, HistGradientBoosting, LightGBM, XGBoost, CatBoost, explainable boosting machines (EBM/GAM), points scorecards and MLPs, matched to data size, width, categorical and high-cardinality share, missingness, class imbalance, text, and constraints such as interpretability, monotonic constraints, calibrated probabilities, latency budgets and tiny deployable models. Ranks candidates with starting hyperparameters and cautions, then generates a bake-off script running every candidate on identical CV folds with per-fold metrics, fit time and prediction latency.
clustering-algorithm-selector
Choose the right clustering algorithm and check that clustering makes sense at all: cluster tendency (Hopkins statistic with a correlated-features caveat, nearest-neighbour isolation), then k-means / MiniBatchKMeans, Gaussian mixtures with BIC, agglomerative (Ward, average, complete, single) with dendrograms, DBSCAN, HDBSCAN, OPTICS, spectral clustering, BIRCH, mean shift, k-prototypes and k-modes for mixed or categorical data, and Gower distance, matched to data size, expected shapes, noise, varying density, known k, soft memberships, hierarchy and streaming. Covers preprocessing (scaling, encoding, dimensionality, distance metrics), validation (internal indices, bootstrap stability, external labels, business profiling) and assigning new points.
collaborative-filtering-implementer
Implement collaborative filtering recommenders that beat popularity: item-based kNN with cosine, Jaccard or asymmetric similarity, shrinkage, neighbourhood size, positive-only or mean-centred rating signals and 'because you liked X' explanations; implicit ALS with confidence weighting, BPR pairwise ranking and the closed-form EASE model in plain NumPy, tuned on a temporal validation split by recall@k and NDCG@k and refit on all data; plus user-kNN, explicit matrix factorisation, the implicit/LightFM libraries, hyperparameter ranges, cold-start fallbacks, scaling with sparse matrices and approximate nearest neighbours. Produces top-k lists in the format of the recommendation evaluator.