HarmonRank: Rank-consistent multi-objective ensembling¶
Core public-data mechanism, not industrial ranking.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Kuaishou Technology. |
| First public release | 2026-01-06 (arXiv v1) |
| Author code | No public author code found as of 2026-09-05. |
| Local adapter or method key | harmonrank |
| Local reproduction code | src/auto_research/reproductions/harmonrank/ |
Paper method and evidence¶
Rewrite AUC as a differentiable rank-sum objective to align tasks in ranking space. Then learn dependencies among objectives such as purchases and comments with self-attention to produce one ranking.
flowchart LR
A[Per-task scores] --> B[Differentiable rank-sum alignment]
B[Differentiable rank-sum alignment] --> C[Objective self-attention]
C[Objective self-attention] --> D[Unified ranking]Original paper figure¶
Figure shows rank-alignment and objective attention. From the original paper; copyright remains with the authors.
Core equation¶
\[
\mathcal L_{rank}=-\sum_k\mathrm{AUC}_{soft}(s_k,y_k),\qquad S=\mathrm{Attn}(R_1,\ldots,R_K).
\]
Reported results¶
Kuaishou live commerce reports purchases +2.635% and watch time +0.451%.
Local reproduction¶
MovieLens-100K fixed full catalog: NDCG@10 0.05401→0.05203 (-3.65%). Three-seed metrics, mean, SD and 95% CI. Auditable proxy, not the production objective. No private features, production traffic or serving. These local public-data results do not reproduce online lift.