HAP: Not All Candidates are Created Equal — Heterogeneity-Aware Pre-ranking¶
Core public-data pre-ranking mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | ByteDance / Toutiao. |
| First public release | 2026-03-04 (arXiv v1) |
| Author code | Toutiao-Rec/HAP. |
| Local adapter or method key | hap |
| Local reproduction code | src/auto_research/reproductions/hap/ |
Paper method and evidence¶
Route candidates to lightweight or strong pre-ranking by difficulty. Cross-branch harmonization keeps ranking scales comparable across compute budgets.
flowchart LR
A[Candidate difficulty] --> B[Route by compute budget]
B[Route by compute budget] --> C[Harmonize branch scores]Original paper figure¶
Figure shows heterogeneous branches and harmonization. From the original paper; copyright remains with the authors.
Core equation¶
\[
s_i=g_i f_{strong}(x_i)+(1-g_i)f_{light}(x_i),\quad\mathcal L=\mathcal L_{rank}+\lambda\mathcal L_{harm}.
\]
Reported results¶
Nine months of Toutiao deployment: usage duration +0.4%, active days +0.05%.
Local reproduction¶
MovieLens-100K full catalog, 260 users/420 items, seed 42: shared transition+content baseline vs HAP 0.0354→0.0371 NDCG@10 (+4.90%).
auto-research reproduce --paper hap --dataset-dir data --seed 42
auto-research evolve --model rankmixer --dataset movielens-100k --direction "Combine hap with installed paper operators" --generations 2 --population 4
Fixed metrics. No private features, production traffic or serving. These local public-data results do not reproduce online lift.