Wide & Deep
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
A neural network design that pairs a wide linear model, which memorizes feature combinations, with a deep network that generalizes to unseen patterns.
What is Wide & Deep?
Wide & Deep is a model architecture that trains a wide linear model and a deep neural network side by side, combining the first’s memorization of exact feature pairs with the second’s generalization to new ones.
Google researchers introduced the design in 2016 for app recommendation on Google Play. It answers a real tension: memorization alone keeps recommending what users already chose, while generalization alone over-recommends plausible-but-wrong items. Wiring both into one jointly trained model gets each to cover the other’s blind spot.
How Wide & Deep Works
- Wide component: A linear model over sparse, crossed features. It memorizes exact co-occurrences, such as “users who installed app A also installed app B”, with no attempt to explain why.
- Deep component: A neural network that maps categorical features through embedding layers and hidden layers. It learns dense representations, so it scores item pairs no user has combined before.
- Joint training: Both components feed one output unit, and their weights update together on the same loss. This differs from an ensemble of separately trained models; each half learns knowing the other exists.
Applications of Wide & Deep
- App and content recommendation: The original Google Play use, ranking items for users with both exact history and inferred taste.
- Search ranking: Blending memorized query-result pairs with generalized relevance signals.
- Ad click-through prediction: Estimating click probability from sparse historical combinations plus dense behavioral features.
Example of Wide & Deep
A food-delivery app ranks restaurants for a user who orders pad thai every Friday. The wide component has memorized that exact pattern: it scores the usual Thai place highly because “this user + this restaurant” appeared many times in training data.
A new Vietnamese restaurant joins the platform with zero order history, so the wide component is silent about it. The deep component, though, has learned embeddings placing Vietnamese and Thai cuisine near each other, and this user near other noodle-soup fans, so it scores the newcomer well.
The combined model surfaces both: the reliable favorite and the plausible discovery. Either component alone would have missed one of them.
Related AI terms: Model Architecture · Neural Network · Embedding Layer · Deep Learning
Did you like the Wide & Deep gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About