DCN-V2¶

Core mechanism: low-rank cross experts, input-dependent gating, and a deep tower.

Paper information¶

Field	Details
Paper	arXiv 2008.13535
Institution	Google.
First public release	2020-08-19 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`dcn-v2`
Local reproduction code	`src/auto_research/reproductions/dcn_v2/`

Paper method and evidence¶

Model bounded-degree feature interactions explicitly. Low-rank factors and a mixture of cross experts reduce ranking parameter and compute costs.

flowchart LR
  X[Input embeddings] --> C[Low-rank cross experts]
  C --> G[Input-dependent mixture]
  X --> D[Deep MLP]
  G --> O[Ranking head]
  D --> O

Original paper figure¶

Figure 2 gives the original cross-layer formulation. From the original paper; copyright remains with the authors.

Core equation¶

\[ x_{l+1}=x_0\odot(W_lx_l+b_l)+x_l,\qquad W_l=\sum_e g_e(x_l)U_eV_e^\top. \]

Reported results¶

The paper reports improvements over DNN/DCN on Criteo, MovieLens, and private Google data, plus a significant business metric improvement without exact online lift. This is a user-approved classic architecture exception, not quantified A/B evidence.

Local reproduction¶

The deep-only baseline is compared with two low-rank cross-expert layers and a gate. Local three-seed NDCG@10 is +22.87% relative. Structured metrics.

auto-research reproduce --paper dcn-v2

No private Google data or production-scale experts; local results only test explicit crossing on public data.