sklearn/missing-random-state
| Rule | missing-random-state |
|---|---|
| Pack | scikit-learn |
| Severity | info |
| Fix | none, edit by hand |
| Library versions | every version |
Without random_state, a split or a model takes a new seed on every run, so the score changes from run to run. Pass random_state.
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)model = RandomForestClassifier()X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0)model = RandomForestClassifier(random_state=0)What the rule reports
Section titled “What the rule reports”The rule reports train_test_split(X, y), RandomForestClassifier() and KFold(shuffle=True)
without random_state.
The rule leaves alone a file that calls np.random.seed, or a seed helper such as set_seed or
seed_everything that the file doesn't define and doesn't import from TensorFlow, PyTorch or
random. It also leaves alone a call with shuffle=False or **, a train_test_split whose
result is thrown away, and a constructor that only isinstance reads. Of 527 held-back reports,
52 were labelled, and each reader called 1 wrong. Recall was 79 of 80.
How to fix it
Section titled “How to fix it”xpd has no automatic fix, so make the change by hand.
Any fixed number makes the run repeatable, and 0 is only the example.
The rule is info, because a seed doesn't move the expected score. A seed can still change which model wins a comparison, as the worse of two models whose true accuracy differs a little won more often without a seed:
| Split | Worse model won, of 300 trials |
|---|---|
| each model on its own split, without a seed | 25.0 percent |
| both models on one shared split | 14.3 percent |
The classes that the rule names gave another output in five fresh runs on scikit-learn 1.9.1:
| Kind | Classes |
|---|---|
| ensembles | the forests, extra trees, bagging, gradient boosting, AdaBoostRegressor, IsolationForest, RandomTreesEmbedding |
| trees and other models | the decision trees, the MLPs, the SGD models |
| clustering | KMeans, MiniBatchKMeans, SpectralClustering, AffinityPropagation |
| decomposition and embedding | TruncatedSVD, NMF, LatentDirichletAllocation, FastICA, MiniBatchDictionaryLearning, MDS, SpectralEmbedding, Nystroem, RBFSampler, the random projections |
| splitters | ShuffleSplit, StratifiedShuffleSplit, RepeatedKFold, RepeatedStratifiedKFold, and KFold, StratifiedKFold, GroupKFold and StratifiedGroupKFold with shuffle=True |
| classes with a setting | TSNE with init="random", SVC(probability=True), a HistGradientBoosting model with early_stopping=True, a Gaussian mixture with more than one component |
A HistGradientBoosting model without random_state is reported with a message that names 10,000
rows, because above that size it stops early on a validation part drawn at random.
LogisticRegression, PCA, Perceptron, AdaBoostClassifier and TSNE with its default init
gave the same output and are not listed.
Message
Section titled “Message”xpd prints this message with the issue. A name that starts with $ stands for code from your file.
train_test_split shuffles with a new seed on every run, so the score changes from run to run. Pass random_state.Browse the other rules of the scikit-learn pack.