What problem does it solve? Poorly chosen encodings, leaky target statistics, and incorrectly scaled features silently degrade model performance or inflate offline metrics that collapse in production. This Skill provides concrete, code-level guidance for building tabular features correctly with scikit-learn and pandas. ## Core Features & Use Cases - Encoding Selection by Cardinality: Decision tables mapping one-hot, ordinal, target, count, hashing, and native categorical encodings to column cardinality and model family, with sklearn code for each. - Leakage Prevention: Cross-fitted, smoothed target encoding via TargetEncoder, point-in-time cutoffs for entity aggregates, and pipeline-based fitting so stateful transforms never see validation data. - Numeric and Temporal Transformation: Scaling guidance per model family, log/quantile transforms for skewed data, cyclical sin/cos time features, and guarded ratio features. - Use Case: When building a fraud model on transaction history, apply the point-in-time aggregate pattern so a user's features never include events after the label timestamp, avoiding a model that detects its own label. ## Quick Start Ask the assistant to apply the feature-engineering skill to design leakage-safe encodings and transformations for your tabular dataset.
