Wide & Deep¶

Fidelity: core mechanism. Combines memorizing wide feature crosses with a generalizing deep tower.

Paper information¶

Field	Details
Paper	arXiv 1606.07792
Organization	Google
First public release	2016-06-24 (arXiv v1)
Authors' code	Public implementation
Adapter / method	`wide-deep`
Local reproduction code	`src/auto_research/reproductions/wide_deep/`

Original paper¶

Motivation and method¶

Linear models memorize sparse co-occurrences but generalize poorly. A pure DNN generalizes but may recommend irrelevant combinations. The paper trains both paths jointly and shares the final logit.

flowchart LR
  A["Sparse features"] --> W["Wide feature crosses"]
  A --> E["Embedding"] --> D["Deep MLP"]
  W --> S["Joint logit"]
  D --> S

Key figure from the paper¶

Figure 4: Core architecture, modules, and their connections. From the original paper; copyright remains with the authors.

Core equation¶

\[ p(y=1\mid x)=\sigma(w_{\rm wide}^{\top}[x,\phi(x)]+w_{\rm deep}^{\top}h_L+b). \]

Paper results¶

In a three-week Google Play online experiment, Wide & Deep increased app acquisitions by 3.9% against the wide control and about 1% against the deep model.

Local reproduction¶

Local comparison: The baseline is a parameter-matched deep-only model; the treatment adds a wide path crossing candidate and genre. Three-seed full-catalog MovieLens-100K NDCG@10 improves +5.45% relative.

auto-research reproduce --paper wide-deep

See the structured results.

Reproduction boundaries¶

Public MovieLens substitutes for private Google Play logs. Production feature engineering, distributed training, and serving were not reproduced.