Deep & Cross Network (DCN)
A deep network fed raw user and item features will, in principle, learn whatever feature combinations matter — but it spends much of its capacity rediscovering something simple: that a user in Mumbai paired with the category cricket gear is a signal in its own right. Deep & Cross Network (DCN) builds those products in explicitly.
Explicit crosses, learned cheaply
DCN scores a candidate with two towers over one shared embedding layer. The cross network stacks layers that each multiply the original input by the current layer's output, so depth k represents interactions up to order k+1 with a parameter count that grows linearly in the input width rather than combinatorially. A conventional MLP runs alongside it for whatever smooth structure remains, and the two outputs are concatenated before the final sigmoid. DCN-V2 replaces the cross layer's weight vector with a low-rank matrix, and is the version production systems use.
Why CTR prediction needs it
Click-through models live on high-cardinality categorical features — device, city, campaign, item category — where the predictive power sits in specific conjunctions rather than any single field. A factorization machine captures those at order two and stops; Wide & Deep captures them at any order, but only the crosses an engineer wrote down by hand. DCN removes that manual step, and that is the operational win: nobody maintains a cross-feature list, and the model still gets the crosses.
Where it sits
DCN is a ranking-stage model: it scores the few hundred candidates retrieval has already narrowed down. Because its score needs the user and the item together, it cannot be precomputed into an embedding index. It sits beside DeepFM and DLRM as a standard baseline on the Criteo benchmark, and ships in TensorFlow Recommenders.
Deep & Cross Network (DCN) is a click-through prediction model that pairs a standard MLP with a cross network — stacked layers that repeatedly multiply the original feature vector by the current output, producing explicit feature interactions of increasing order at linear parameter cost. It yields Wide & Deep's hand-written cross features automatically, which matters for the high-cardinality categorical fields CTR ranking depends on. It is used at the ranking stage, not for retrieval.
When a service guesses whether you will click something, the useful clues are rarely single facts. Lives in a cold country is weak on its own. Lives in a cold country, and browsing winter coats, and it is November — together that is strong. Combinations carry the signal.
Deep & Cross Network is a design that makes a prediction model find those combinations by itself. Older systems worked, but only because engineers sat down and listed the pairings they thought mattered, a slow job that missed things nobody thought of. DCN multiplies the available facts together layer by layer, so useful combinations surface on their own, and the model stays small enough to answer in the moment the page loads.