Classification Sktime

sktime Classification

Use this skill after ts-classification-data-prep. sktime is best when the task is supervised assignment of complete time-series instances to predefined classes, using sklearn-like fit, predict, predict_proba, pipelines, cross-validation, and a broad zoo of time-series classifiers.

Do not use forecasting APIs for classification. Each row/instance is a whole time series sample; labels are class labels for instances, not future values.

Minimum Install

pip install sktime
pip install "sktime[all_extras]"

Start with core sktime; install soft dependencies only when an estimator asks for them. Deep learning, pyts, tslearn, torch, or other adapters may require extras or separate packages.

Data Contract

• Require a completed ts-classification-data-prep contract: X, y, class labels, split IDs, instance order, channel names, sampling policy, padding/truncation policy, missing-value policy, and leakage notes.
• Preferred Panel formats: numpy3D with shape (n_instances, n_variables, n_timepoints) or pd-multiindex with row index (instance, time) and columns as variables.
• df-list can represent variable length, unequal indexes, and different variable sets; pd-multiindex can represent unequal support but not different variable sets; numpy3D requires equal length/index and the same variables.
• y is a 1D array-like label vector with one label per instance.
• Use sktime.datatypes.check_raise/check_is_mtype or scripts/validate_numpy3d.py before fitting.

Read references/sktime-data-validation.md before using variable length, multivariate data, missing values, padding/truncation, sklearn pipelines, or CV.

Classifier Selection

Use official class names from sktime.classification.*. Main families:

• Distance/kernel: KNeighborsTimeSeriesClassifier, ElasticEnsemble, ShapeDTW, proximity tree/forest, TimeSeriesSVC.
• Dictionary: BOSSEnsemble, ContractableBOSS, IndividualBOSS, IndividualTDE, MUSE, TemporalDictionaryEnsemble, WEASEL, MrSEQL, MrSQM.
• Interval/forest: TimeSeriesForestClassifier, CanonicalIntervalForest, DrCIF, RandomIntervalSpectralEnsemble, SupervisedTimeSeriesForest.
• Feature extraction: Catch22Classifier, SummaryClassifier, TSFreshClassifier, FreshPRINCE, RandomIntervalClassifier, SignatureClassifier, MatrixProfileClassifier.
• Shapelet/ROCKET/hybrid: ShapeletTransformClassifier, shapelet-learning adapters, RocketClassifier, Arsenal, HIVECOTEV1, HIVECOTEV2.
• Deep/foundation: CNN/FCN/ResNet/RNN/GRU/LSTM/InceptionTime/TapNet/MVTSTransformer variants plus MantisClassifier, MomentFMClassifier, TSPulseClassifier.
• Composition: ClassifierPipeline, SklearnClassifierPipeline, ColumnEnsembleClassifier, MultiplexClassifier, BaggingClassifier, WeightedEnsembleClassifier, TSCGridSearchCV.

Read references/sktime-classifier-map.md before claiming a classifier, dependency, multivariate support, unequal-length support, or predict_proba support.

Training Pattern

from sklearn.metrics import classification_report
from sklearn.model_selection import StratifiedKFold, cross_validate
from sktime.classification.feature_based import SummaryClassifier

clf = SummaryClassifier()
clf.fit(X_train, y_train)

y_pred = clf.predict(X_test)
y_proba = clf.predict_proba(X_test) if clf.get_tag("capability:predict_proba") else None

cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
scores = cross_validate(
    clf,
    X_train,
    y_train,
    cv=cv,
    scoring=["accuracy", "f1_macro", "balanced_accuracy"],
)
print(classification_report(y_test, y_pred))

Use ClassifierPipeline or make_pipeline so transformations are fit inside each train fold:

from sktime.classification.distance_based import KNeighborsTimeSeriesClassifier
from sktime.transformations.exponent import ExponentTransformer

pipe = ExponentTransformer() * KNeighborsTimeSeriesClassifier()

Feature Extraction And Shape Handling

• Use classifier-native feature extractors where possible, e.g. Catch22Classifier, SummaryClassifier, TSFreshClassifier, ROCKET/shapelet classifiers.
• If using standalone transforms such as Rocket, MiniRocket, ShapeletTransform, RandomShapeletTransform, SummaryTransformer, Catch22, or TSFreshFeatureExtractor, put them in a pipeline before a classifier.
• For unequal length, either select classifiers/tags with capability:unequal_length=True or add a train-fitted transformer such as PaddingTransformer or TruncationTransformer inside the pipeline.
• For missing values, use Imputer or a classifier that documents missing-value capability.

Evaluation

• Use stratified splits for class imbalance: train_test_split(..., stratify=y) and StratifiedKFold.
• Balanced classes: report accuracy plus macro/weighted F1.
• Imbalanced classes: prefer balanced_accuracy, f1_macro, per-class recall/precision, confusion matrix, and ROC-AUC/average precision when probabilities and problem shape support them.
• Use TSCGridSearchCV or sklearn GridSearchCV/RandomizedSearchCV; keep every scaler, padder, shapelet/dictionary/ROCKET feature extractor, and model inside the CV pipeline.

Anti-Leakage Rules

• Split into train/validation/test before fitting any scaler, padder, imputer, shapelet, dictionary, ROCKET, TSFresh, PCA, feature selector, or resampler.
• Fit transformations only on train folds via ClassifierPipeline, SklearnClassifierPipeline, or sklearn/sktime CV wrappers.
• Do not compute global padding/truncation lengths from test data unless that length is a fixed external contract from ts-classification-data-prep.
• Use stratified CV for imbalanced labels.
• Keep grouped subjects/devices out of both train and test if the prep contract identifies grouping; use group-aware stratified splitting where needed.

Common Errors

• Passing (n_samples, series_length, n_channels) instead of sktime numpy3D (n_samples, n_channels, series_length).
• Using a classifier whose tags do not support multivariate, unequal length, missing values, or predict_proba.
• Running feature extraction on the full dataset before CV.
• Using a plain sklearn classifier directly on numpy3D; wrap feature extraction with SklearnClassifierPipeline.
• Assuming all API-listed deep/foundation classifiers are available without soft dependencies.
• Using non-stratified CV on heavily imbalanced labels.

References

• Read references/sktime-classifier-map.md for classifier families, official class names, tags, and dependency caveats.
• Read references/sktime-data-validation.md for accepted mtypes, padding/truncation, cross-validation, metrics, and leakage-safe pipelines.
• Read references/official-sources.md for official sources consulted.
• Use scripts/validate_numpy3d.py to validate .npy X and .npy/CSV labels before fitting.

Ready Checklist

• ts-classification-data-prep contract is complete and leakage risks are documented.
• X is a valid sktime Panel mtype and len(y) == n_instances.
• Equal/variable length, multivariate channels, missing values, and class imbalance are explicitly handled.
• All transformations and feature extraction live inside the fitted/CV pipeline.
• Classifier tags match data requirements and predict_proba expectations.
• Evaluation uses stratified temporal/group-aware split if required by the task, and reports imbalance-aware metrics.