data-feature

To prevent target leakage during feature engineering, enforce strict train/test split rules and perform leakage checks before applying any transformations like lags or rolling features. This validation ensures future data does not incorrectly influence predictive modeling.

The best way to generate predictive features is mapping domain hypotheses to concrete candidate features, then applying transformations like encoding, scaling, and time-series lags. This process outputs an auditable feature engineering report for downstream modeling.

You check multicollinearity by running Variance Inflation Factor (VIF) checks on your candidate features. Performing these checks before applying transformations helps identify and flag redundant predictive features early in the workflow.

Yes, you can estimate feature importance for tabular datasets using SHAP or permutation importance methods. This validates the predictive value of generated features after applying necessary encodings and safety checks.

Feature engineering can technically work without it, but this process specifically reads analysis_context.md as the single source of truth within a CRISP-DM workflow to propose candidate features and document the execution log.