Preprocessing placement and leakage

Use this guide to decide where a preprocessing node belongs relative to the Train-Test Split node. It covers 62 registered preprocessing IDs: 60 transformations and two row-split registrations. Aliases have separate entries because older saved pipelines can contain either spelling.

Placement depends on the configured operation, not just the node's display name. The same node can apply a fixed formula in one configuration and learn a statistic in another. These rules describe supported implementations and parameter modes; they do not guarantee that an arbitrary dataset or experiment is leakage-free.

Start with the training boundary

A practical linear pipeline is:

Loader -> Fixed rules -> Train-Test Split -> Learned preprocessing -> Train / Tune

Learned preprocessing must fit using training rows only. Apply the resulting state to validation, test and inference rows without fitting it again. Examples include scaling statistics, imputation values, category vocabularies, feature selection, bin edges, target statistics and group lookup tables.

Cross-validation introduces an inner boundary. Fit learned preprocessing again inside each training fold, then apply it to that fold's validation rows. A scaler fitted once on the whole outer training set has already seen the inner validation folds. Putting it after the outer split alone does not resolve that problem.

Resampling is stricter: oversampling and undersampling act only on training rows, including the training partition of each CV fold. Do not resample the validation or test population.

Feature-Target Split is not Train-Test Split. Separating X from y does not reserve any rows for evaluation. TrainTestSplitter and its legacy Split alias create the row boundary checked by the leakage gate.

Prediction row alignment

SkyulfPipeline.predict() and deployment prediction require one result per submitted row. If an applied step filters or adds rows, prediction stops with the step name and before/after counts. For example, an IQR step that reduces [2, 1000, 4] to two rows cannot return an ambiguous two-element prediction list. Use row-preserving handling such as Winsorize, or filter the raw input yourself and retain its row mapping before submitting it.

LagFeatures and RollingAggregate can reorder rows through sort_by. Prediction also rejects a batch whose order changes in either built-in step; sort the input as required before submitting it. Temporary position checks stay outside feature columns and are removed before the next step. Custom transformers remain responsible for preserving order; the generic guard checks their row counts.

Ordinary FeatureEngineer.transform() and CV/threshold scoring retain configured filtering and sorting. The prediction guard uses preserve_rows=True and does not change training, fitted artifacts or validation-label alignment. Existing splitter, resampling, Deduplicate and DropMissingRows inference skips remain.

Read the placement labels

Label	Meaning	What to do
Before split (fixed)	Uses configured rules, schema or a pretrained model without fitting the current dataset.	May precede the split when its inputs and configuration are valid for prediction.
Depends on operation	Some configurations are fixed; others learn from data.	Inspect every configured operation and selection rule.
After split	Active preprocessing learns state or is conservatively gated.	Fit on training rows and refit inside CV training folds. Resampling is training-only.
Time/history review	Uses neighboring or earlier rows without fitting a distribution.	Check ordering, entity groups, horizon, history availability and split strategy.
Creates row split	Establishes the train/test row boundary.	Place learned feature preprocessing downstream and choose an appropriate split protocol.

"Before split" is permission to apply a fixed operation, not a universal safety claim. A ratio that includes the answer, a timestamp recorded after an outcome, or a manual threshold chosen from test results can leak even though the operation fits no statistic. Reporting statistics is not feature fitting, but using test reports to select features or tune a model contaminates the evaluation.

Configuration changes that change placement

General Transformation and Power Transformer

Fixed logarithms, square roots and squares do not fit a parameter from the dataset. Yeo-Johnson and Box-Cox do. Put those learned methods after the row split and fit them within each CV training fold. Turning off standardization does not make the power parameter fixed.

Inspect the entire transformation list. If one entry learns, the mixed node learns. The name "General Transformation" alone cannot establish its placement.

Polynomial Features and its alias

PolynomialFeatures and PolynomialFeaturesNode build a fixed mathematical basis, but their optional automatic column selection learns from the fitting rows.

Configuration	Placement
auto_detect: true, with columns omitted or []	After the row split: discover eligible columns from training data only.
Explicit nonempty columns, even with auto_detect: true	Before split is allowed: column selection and polynomial math are fixed.
auto_detect omitted or false	Before split is allowed: automatic discovery is disabled.

The registry conservatively marks both aliases as learned; the configured operation determines placement. FeatureInteraction is a separate fixed node and does not inherit this automatic-selection rule.

Feature Generation and its aliases

Arithmetic, ratios, similarity and datetime extraction are fixed calculations. group_agg fits a group lookup on training rows and reuses it on held-out rows. Unknown groups become missing instead of obtaining statistics from the test batch; add suitable downstream missing-value handling when needed.

This distinction applies to FeatureGeneration, FeatureGenerationNode and FeatureMath. A mixed operation list containing a learned group aggregate belongs after the split, even when its other entries are arithmetic. Fixed expressions still require inputs that exist at prediction time.

Casting, imputation and column deletion

Configuration	Placement reason
Casting numeric/string/datetime values	Fixed conversion.
Casting to category/categorical	Fits a training vocabulary; unseen categories become missing. Resolve per-column overrides first.
Simple Imputer with a configured constant	Fixed fill rule.
Simple Imputer with mean, median or most-frequent strategy	Fits training statistics.
Drop Missing Columns without a positive threshold	Fixed configured deletion.
Drop Missing Columns with positive missing_threshold	Learns which columns to remove, even with explicit candidate columns.

Column selection is part of the operation

Do not treat omitted columns, null and [] as interchangeable:

Node	Omitted/null selection	Explicit empty selection	Explicit nonempty selection
Custom Binning	Value-based column discovery; learned.	Fixed/no-op.	Configured columns and edges are fixed.
Hash Encoder	Automatic selection; conservatively gated.	Fixed/no-op.	Fixed hashing of selected columns.
Missing Indicator	Discovers columns containing missing values; learned.	Also automatic discovery; learned.	Fixed indicator columns.
Label Encoder	Available target only, or no-op.	Available target only, or no-op.	Feature columns learn mappings; exact known-target selection is exempt.
Ordinal Encoder	Discovers and learns feature categories.	Target-only/no-op.	Feature encoding remains learned, including configured category orders; exact known-target selection is exempt.
Count and TF-IDF Vectorizers	No-op; no automatic text discovery.	No-op.	Feature text learns vocabulary and, for TF-IDF, IDF.
Hashing Vectorizer, Tokenizer, Sentence Embedder	No-op.	No-op.	Fixed processing or pretrained embeddings.
Text Cleaning	Discovers text dtypes for fixed cleaning.	No-op.	Fixed cleaning of selected text columns.

Target-only exemptions require an authoritative target in the current pipeline or branch. A target name from an unrelated branch is not evidence that an encoder is operating only on labels. Text feature processing excludes the authoritative target.

An explicit [] is also a supported no-op for OneHotEncoder, DummyEncoder, TargetEncoder, WOEEncoder, PowerTransformer, StandardScaler, MinMaxScaler, MaxAbsScaler, RobustScaler, SimpleImputer, KNNImputer, IterativeImputer, GeneralBinning, KBinsDiscretizer, CustomBinning, IQR, ZScore, Winsorize and EllipticEnvelope. This is a specific implementation contract, not a rule to apply to other nodes. Selectors, resamplers and Missing Indicator do not receive a general empty-list exemption.

Target encoders, lags and rolling windows

TargetEncoder learns target statistics. Training rows in the pipeline receive out-of-fold encodings; held-out and inference rows use the full-training artifact. Unknown categories fall back to the training prior. WOEEncoder uses complement-fold mappings for training rows and the full-training mapping for held-out rows, with the configured unknown-category fallback, currently zero. These protections do not remove the need for the outer row boundary and CV preprocessing refits.

WOE smoothing adds regularization (default 0.5) to every observed category's count in each target class. Each class total therefore includes regularization * number_of_observed_categories; missing feature values count as a category. Cross-fitting uses only the categories in each training complement. The corrected WOE and Information Value calculations apply to new fits; existing saved mappings retain their values. Refit the pipeline to use the correction.

LagFeatures and RollingAggregate do not fit a distribution, but they depend on other rows. Use the correct time ordering, entity groups, prediction horizon and history protocol. A rolling aggregate includes the current row: a current-target input can expose the answer directly. A lagged target is useful only when that target observation is actually available at the intended prediction time. Merely moving a temporal node after a random split does not establish a valid protocol.

Newly fitted DateFeatures steps extract calendar parts in UTC. For example, 2024-01-01 00:30+02:00 produces year 2023, day 31 and hour 22. Naive dates are interpreted as UTC without moving their clock time; mixed offsets and DST transitions follow the same rule on both engines. The original column remains unchanged unless drop_original is enabled. Saved older artifacts retain their original engine-specific behavior so an existing model's features do not change on reload. Refit the complete pipeline to adopt UTC for those models.

Complete registered-node catalog

The placement of learned entries below describes active feature processing. The explicit-empty and target-only exceptions above still apply. Aliases remain separate.

Registered ID	Placement	Behavior and practical rule
AliasReplacement	Before split (fixed)	Apply configured/domain aliases to each value; no fitted vocabulary.
Casting	Depends on operation	Fixed casts may precede split; categorical vocabularies fit on train. Resolve per-column overrides.
CorrelationThreshold	After split	Learn retained columns from training correlations; empty selection still invokes discovery.
CustomBinning	Depends on operation	Explicit columns and configured edges are fixed; omitted/null columns learn value-based selection.
DataSnapshot	Before split (fixed)	Report a snapshot without changing modeling data; do not tune from held-out reports.
DatasetProfile	Before split (fixed)	Report statistics while preserving data; held-out-driven modeling decisions can still leak.
DateFeatures	Before split (fixed)	Extract configured calendar fields from each timestamp.
Deduplicate	After split	Conservatively gated cross-row filtering. Separately prevent duplicate entities from crossing evaluation splits.
DropMissingColumns	Depends on operation	Positive missing threshold learns deletion; otherwise use fixed configured columns.
DropMissingRows	Before split (fixed)	Apply fixed per-row missingness rules. Removing evaluation rows changes the scored population.
DummyEncoder	After split	Learn training categories and fixed indicator schema; unknown categories yield zeros.
EllipticEnvelope	After split	Fit training outlier estimator and reuse it. Filtering held-out rows changes evaluation coverage.
FeatureGeneration	Depends on operation	Fixed arithmetic/ratio/similarity/datetime; group_agg learns training lookup.
FeatureGenerationNode	Depends on operation	Feature-generation alias with the same mixed-operation rules.
FeatureInteraction	Before split (fixed)	Build a polynomial basis from schema and configured options.
FeatureMath	Depends on operation	Feature-generation alias; any learned group aggregate makes the node learned.
GeneralBinning	After split	Learn bin edges or discretization state on training values and replay them.
GeneralTransformation	Depends on operation	Fixed math may precede split; Yeo-Johnson/Box-Cox fit parameters. Inspect all rules.
GeoDistance	Before split (fixed)	Compute configured distances from each row's coordinates.
H3Index	Before split (fixed)	Map coordinates to cells at a configured resolution.
HashEncoder	Depends on operation	Explicit columns use fixed hashing; automatic selection is conservatively gated.
IQR	After split	Learn training quartiles and bounds; held-out outliers cannot move them.
InvalidValueReplacement	Before split (fixed)	Apply configured invalid-value replacement rules.
IterativeImputer	After split	Fit imputation models on training data and reuse them.
KBinsDiscretizer	After split	Binning alias; active data-derived modes learn training edges/state.
KNNImputer	After split	Fit/store training reference rows; do not use held-out rows as fitting references.
LabelEncoder	Depends on operation	Default/empty target-only or no-op; explicit feature columns learn category mappings.
LagFeatures	Time/history review	Positive lags require valid ordering, groups, available history and prediction horizon.
ManualBounds	Before split (fixed)	Apply supplied bounds; do not choose them by inspecting test outcomes.
MaxAbsScaler	After split	Fit training maximum absolute values and reuse the scale.
MinMaxScaler	After split	Fit training extrema; held-out extrema cannot redefine the range.
MissingIndicator	Depends on operation	Nonempty explicit columns are fixed; omitted/null/[] learns missing-column discovery.
ModelBasedSelection	After split	Fit selector/model on training features and targets as required.
OneHotEncoder	After split	Fit categories/frequency grouping and preserve the training output schema.
OrdinalEncoder	Depends on operation	Default learns feature categories; explicit [] or exact known-target selection is exempt.
Oversampling	After split, training only	Resample each training partition only, never validation or test rows.
PolynomialFeatures	Depends on operation	auto_detect: true with omitted/empty columns learns column selection; otherwise fixed polynomial math.
PolynomialFeaturesNode	Depends on operation	Same automatic-selection rule as PolynomialFeatures; explicit nonempty columns remain fixed.
PowerTransformer	After split	Fit Yeo-Johnson/Box-Cox parameters even when standardization is disabled.
RobustScaler	After split	Fit robust training location and scale statistics.
RollingAggregate	Time/history review	Ordered/grouped rolling features include the current row; current-target input can leak directly.
SimpleImputer	Depends on operation	Configured constant is fixed; mean/median/most-frequent fits training statistics.
SimpleTransformation	Before split (fixed)	Apply configured simple math without fitting power parameters.
Split	Creates row split	Legacy alias of TrainTestSplitter.
StandardScaler	After split	Fit training centering/scaling statistics and refit within CV folds.
TargetEncoder	After split	OOF training encoding; full-training artifact for held-out/inference, training prior for unknowns.
TextCleaning	Before split (fixed)	Fixed per-value cleaning; omitted/null detects text dtypes, [] does nothing.
TrainTestSplitter	Creates row split	Create the actual train/test row boundary with a suitable split protocol.
Undersampling	After split, training only	Preserve the original validation and test populations.
UnivariateSelection	After split	Fit feature scores/selection on training rows and targets as required.
ValueReplacement	Before split (fixed)	Apply explicitly configured value replacements.
VarianceThreshold	After split	Fit feature variances and selection; [] is not a general no-op exemption.
WOEEncoder	After split	Complement-fold training mappings; full-training held-out mapping and configured unknown fallback.
Winsorize	After split	Fit training clipping limits; held-out extreme values cannot change them.
ZScore	After split	Fit training location/scale and reuse the stored outlier rule.
count_vectorizer	Depends on operation	Nonempty feature selection learns vocabulary; omitted/null/[] is no-op.
feature_selection	After split	Dispatch to configured learned selector; preserve training-only and fold-refit boundaries.
feature_target_split	Before split (fixed)	Separate X from y; this is not a train/test row boundary.
hashing_vectorizer	Before split (fixed)	Fixed hash space; explicit nonempty text columns required.
sentence_embedder	Before split (fixed)	Apply pretrained embeddings; explicit nonempty columns required, external model provenance remains separate.
tfidf_vectorizer	Depends on operation	Nonempty feature selection learns vocabulary/IDF; omitted/null/[] is no-op.
tokenizer	Before split (fixed)	Fixed tokenization; explicit nonempty text columns required.

Supported graph shapes and current policy

Prefer a linear chain when the feature operations are sequential. Backend fold reconstruction also supports these branch forms when their requirements hold:

• A common row splitter forks into linear transformer branches that preserve row alignment and join directly into the model. Avoid additional splitters, nested merges and disallowed row-changing steps in the branches.
• Genuinely fixed branches from a shared loader merge at the first row splitter, followed by a reconstructable downstream chain. The resolver retains merged raw input and outer training row selection for fold-local fitting.

Supported topology does not establish correct feature ownership. Preserve intended column selection and merge order; overlapping outputs can replace earlier branch columns. A join into an intermediate encoder is not the same supported form as a direct join into the model. See the core/backend guide for diagrams and exact reconstruction requirements.

With the default raise policy, a detected definite preprocessing-placement violation is rejected by the backend admission gate with HTTP 400 and no jobs created. Unsupported fold reconstruction involving learned preprocessing also stops execution by default, but that capability failure can occur after a job has been created. These are separate outcomes.

Explicit metadata.on_leakage="warn" or "ignore" can permit the legacy fallback to pre-transformed data. Its CV scores can be optimistically biased, and its fold_refit_fallback diagnostic describes the missing guarantee. Choosing a weaker policy does not make the graph reconstructable or repair the statistical problem.

Read the Leakage Gate and Fold Refit Audit separately. Green results provide evidence for the checks performed, not proof of feature provenance, entity independence, valid temporal assumptions or an untouched test-selection process.

Core-only CV callers must fit a fresh pipeline on each fold's raw training partition; cross-validating an estimator over a globally preprocessed matrix does not automatically refit the core preprocessing steps.

Where to use this reference

The canvas How pipelines work dialog contains a searchable Preprocessing & Leakage tab with the same registered-node coverage. Search by display name, registry ID, category or operation, and filter by placement.

For implementation evidence and the exact fit/apply contracts, see the preprocessing leakage audit. For graph diagrams, admission and fold-reconstruction details, see Leakage: core and backend.