Feature Crosses & Wide-and-Deep

Wide-and-deep (the Google Play paper) splits the ranker into two personalities:

• Wide — a linear model over hand-built feature crosses. It memorizes: 'users who installed French-vocab apps also installed recipe apps' becomes a specific weighted rule.
• Deep — embeddings plus hidden layers. It generalizes: it can score a (user, app) pair nobody has ever seen because it interpolates through the embedding space.

Memorization alone can't extrapolate; generalization alone misses the weird, specific, high-value rules that drive real engagement. Production rankers bet on both.