简体中文 · English

HA-MoE: Unified industrial ranking for heterogeneous content¶

Core mechanism reproduction. Heterogeneity measurements, specialized experts and multi-gate fusion run locally. MovieLens is a proxy, not Google Discover production.

Paper information¶

Field	Details
Paper	arXiv 2607.27577
Company	Google / Discover
First public release	2026-07-30 (arXiv v1)
Author code	No official author code in the paper when checked on 2026-08-09.
Adapter	`ha-moe`
Local reproduction code	`src/auto_research/reproductions/ha_moe/`

Original paper summary¶

Background and main changes¶

Discover ranks news, video, sports and other open-web content together. With one shared tower, frequent content can dominate gradients. HA-MoE controls multi-gate expert mixing with content heterogeneity to separate common signals from domain expertise. LENS tracks expert usage and per-domain DL-AUC so aggregate scores do not hide local regressions.

flowchart LR
  X["User, content, and context"] --> E["Shared backbone"]
  H["Content type and heterogeneity"] --> G["Heterogeneity-aware gates"]
  E --> M["Domain experts"]
  G --> M
  M --> R["Unified ranking score"]
  R --> L["LENS / DL-AUC monitoring"]

Key figure from the paper¶

Figure 2: HA-MoE and unified ranking across heterogeneous content. From the paper; copyright belongs to the authors.

Core equation¶

For input \(x\), the gating network weights experts by heterogeneous context \(h\):

\[ y(x,h)=\sum_{e=1}^{E}\operatorname{softmax}(g(h))_e f_e(x). \]

The local context \(h\) comes from session domain entropy. Domain, transition, content and freshness experts all run.

Offline and online results¶

Seven days of A/B on 1% live traffic: DAU +0.22%±0.11%, Viewed Impressions +0.48%±0.34%.
Scroll Depth +0.34%±0.25%, Diverse Feed Rate +0.36%±0.03%, Diverse Engagement Rate +0.54%±0.07%.

Local reproduction¶

Comparison: Homogeneous transition-content ranker versus HA-MoE on the same split. Relative NDCG@10: -1.77%.

Full-catalog ranking on fixed MovieLens 100K: 220 users / 360 items. Use the same transition-content control and choose blend only on validation. Hit@10 is tied, NDCG@10 is -1.77%, and head share is -7.65%. Diversity improves on this small dataset without a primary-metric gain. The negative result is retained.

See metrics/movielens-100k-seed42.json.

auto-research reproduce --paper ha-moe --dataset-dir data --seed 42

Scope¶

MovieLens genres proxy open-web content types. Discover private samples, production LENS monitoring and DL-AUC services are not reproduced.