RankUp: High-rank representations for advertising recommendation¶
Core public-data representation mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Tencent. |
| First public release | 2026-04-20 (arXiv v1) |
| Author code | No public author code found as of 2026-09-05. |
| Local adapter or method key | rankup |
| Local reproduction code | src/auto_research/reproductions/rankup/ |
Paper method and evidence¶
Counter representation collapse with randomized grouping, multiple embeddings, global tokens, cross-pretrained embedding injection and decoupled task tokens so added capacity increases representational rank.
flowchart LR
A[Sparse embeddings] --> B[Permute and regroup]
B[Permute and regroup] --> C[Inject pretrained and global tokens]
C[Inject pretrained and global tokens] --> D[Separate task tokens]Original paper figure¶
Figure shows the high-rank feature pipeline. From the original paper; copyright remains with the authors.
Core equation¶
\[
H=\operatorname{Concat}_{m=1}^{M}\phi\!\left(E_m(XP_m)W_m\right),\qquad H'=H+g.
\]
Reported results¶
Full deployment on WeChat Channels, Official Accounts and Moments reports GMV +3.41%, +4.81%, +2.12%, respectively.
Local reproduction¶
MovieLens-1M fixed catalog: NDCG@10 0.05401→0.04702 (-12.94%). Three-seed metrics, mean, SD and 95% CI. No private features, production traffic or serving. These local public-data results do not reproduce online lift.