What are feature crosses? When would you explicitly create them instead of relying on a model to learn interactions?
formulate your answer, then —
tldr
Feature crosses let models capture interactions, especially in linear or sparse categorical systems. They improve memorization but can explode cardinality and overfit. Use them with support thresholds, hashing, regularization, and careful train-serve consistency.
follow-up
- Why do linear models need explicit feature crosses?
- When would embeddings be better than crossed one-hot features?
- How do feature crosses create train-serve skew?