简体中文 · English
RAMP: Robust ad recommendation with restricted personalization features¶
Core mechanism reproduction. Masked dual pathways and prediction alignment are actually trained. Private industrial CVR data is not reproduced.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2607.17473 |
| Institutions | Huawei Ireland Research Center / University College Dublin |
| First public release | 2026-07-20 (arXiv v1) |
| Author code | Yes: official repository |
| Adapter | ramp |
| Local reproduction code | src/auto_research/reproductions/ramp/ |
Original paper summary¶
Background and main changes¶
Privacy choices, permissions and platform limits can hide personalized fields for some traffic. Instead of zero-filling, RAMP maintains personalized and non-personalized paths, selects output with an availability mask, and aligns predictions to distill knowledge from the feature-rich path into the restricted one.
flowchart LR
A["Public ad fields"] --> B["Non-personalized pathway"]
C["Personalized user fields"] --> D["Personalized pathway"]
E["Feature availability mask"] --> F["Output selection"]
B --> F
D --> F
D -. "KL alignment" .-> BKey figure from the paper¶
Figure 2: Core architecture and connections. From the paper; copyright belongs to the authors. Click for the source.
Core equations¶
Offline and online results¶
Relative AUC gains on the paper's public data are around 0.10%–0.87%, with offline significance \(p<0.001\). Industrial CVR A/B reports more than 3% lift in Total Advertiser Value.
Local reproduction¶
MovieLens behavior vectors proxy restricted personalization fields, and genres proxy public ad fields. Training runs both paths, output masking and KL alignment from the feature-rich teacher to the restricted path. Test specifically evaluates traffic without personalized fields.
Comparison: Shared CTR tower with missing personalization versus the RAMP masked dual tower. Seed 42 NDCG@10 rises from 0.00103 to 0.00533, +417.23% relative.
Stable metrics: metrics/movielens-100k-seed42.json. Absolute values are low and the proxy task is much smaller than industrial CVR. This shows relative robustness to missing features only.