Recursive feature elimination (RFE) selects features iteratively: fit a model, rank features by importance, remove the worst K, and repeat. It offers model-aware selection but can be computationally expensive, so it is usually better suited to deliberate refinement than the first rough pruning pass. section 11.3.3
Use RFE as part of a practical feature-selection workflow:
- Start with rough cutoffs when many features are obviously irrelevant; reserve RFE for more meaningful subset decisions. section 11.3.3
- Use an embedded model’s feature importances to decide which features to remove at each iteration. section 11.3.3
- Check the selected features against model-agnostic importance measures, feature-quality tests, correlation limits, and computation-time constraints. section 11.5.1
Demand forecasting at Supermegaretail
During early feature development, RFE automatically reduces the candidate set for SKU-store-day demand models. The team then tests ranges, outliers, pairwise correlation, contribution, and feature computation time before and after training. section 11.5.1
