Predicting whether someone will click an ad depends heavily on combinations of features, such as a particular user segment seeing a particular product category on a particular device. Capturing these “feature crosses” traditionally meant hand-engineering them, which is laborious and incomplete. In 2017 Ruoxi Wang and colleagues at Google and Stanford presented the Deep & Cross Network (DCN) at the AdKDD workshop to learn such crosses automatically.
DCN keeps the deep neural network familiar from Wide & Deep but replaces the hand-built wide part with a “cross network.” The cross network applies an explicit feature-crossing operation at each layer, so stacking layers produces higher-degree feature interactions in a controlled, bounded way, with very little added computation and no manual feature engineering. The deep and cross components run in parallel and their outputs are combined for the final prediction.
The design improved click-through-rate prediction accuracy over strong baselines while staying efficient, and a later revision, DCN-V2, made the cross layers more expressive for industrial use. The model became a common building block in advertising and recommendation ranking stacks.
For a business reader, DCN is part of the quiet machinery behind online advertising: it is one of the ways platforms automatically discover which combinations of signals predict a click, which is what turns attention into revenue.