DCN-v2 Deep & Cross Network

DCN-v2 (Deep & Cross Network V2, proposed by Google in 2021) is a high-performance ranking network that explicitly and efficiently models arbitrary bounded-degree polynomial feature interactions; its core Cross Network evolves via layer-wise recursion: $x_{l+1} = x_0 \odot (W_l x_l + b_l) + x_l$, where $\odot$ denotes the Hadamard product and each subsequent layer increments interaction degree by 1 ($l$ layers model $(l+1)$-th order interactions); to control compute over massive feature dimensions, DCN-v2 introduces Low-Rank Matrix Factorization ($W = U V^T, r \ll d$) and Mixture-of-Experts gating.

Heavy ranking in production recommendation systems, sponsored search CTR/CVR estimation over high-dimensional sparse inputs.

Standard MLPs learn feature interactions implicitly and inefficiently; DCN-v1 was bottlenecked by rank-1 parameter limits; DCN-v2 unlocks full-matrix and low-rank explicit polynomial interactions, delivering state-of-the-art accuracy within production inference budgets.

Derive how the Cross Network recursion $x_{l+1} = x_0 \odot (W_l x_l + b_l) + x_l$ explicitly increments interaction degrees layer-by-layer?

How does DCN-v2's Low-Rank factorization ($W = U V^T$) slash computational complexity from $O(d^2)$ to $O(2dr)$?

Compare Parallel topology vs Stacked topology in DCN-v2 in terms of gradient propagation and feature synergy?

How does the MoE mechanism in DCN-v2 assign inputs across specialized expert sub-networks to capture multimodal interactions?

Benchmark floating point operations (FLOPs) and inference throughput (QPS) of DCN-v2 Cross Networks against traditional deep MLPs?