RAMP: New AI system for ad recommendations that works when user data is missing
CTR predictions stay accurate even when age and gender data are blocked by privacy rules
Click-through rate (CTR) and conversion rate (CVR) prediction rely heavily on personalized user features—like age and gender—to serve relevant ads. But tightening privacy regulations (e.g., GDPR, CCPA) increasingly restrict access to this data, hurting model accuracy. Enter RAMP (Robust Ad Recommendation under Limited Personalized-Feature Availability via Masking and Alignment Pathways), a new framework designed to keep ad recommendation systems performing well even when those features are missing.
RAMP’s architecture has three components: (1) a personalized pathway built on a dual-tower model with identical inputs but independent parameters, where output masking separates predictions for personalized and non-personalized signals; (2) a separate non-personalized pathway trained solely on non-personalized features; and (3) a distillation-inspired prediction-alignment mechanism that bridges the two pathways. This design lets the model learn robust representations and maintain accuracy when personalized features vanish during inference.
In experiments across multiple backbone models, public benchmarks, and industrial datasets, RAMP consistently outperformed state-of-the-art methods when personalized features were unavailable—while matching or exceeding baseline performance when all features were present. The paper (12 pages, 4 figures) has been accepted at ICTIR '26, and the code is publicly available on GitHub. For advertisers and platforms facing privacy restrictions, RAMP offers a practical way to preserve ad relevance without compromising user privacy.
- RAMP uses a dual-tower architecture with output masking to separate personalized and non-personalized signals, plus a prediction-alignment loss to ensure robustness when personalized features are missing.
- Tested on both public benchmarks and industrial datasets, RAMP consistently outperforms state-of-the-art CTR/CVR models under privacy-constrained settings.
- Code is open-source; the paper was accepted at ICTIR '26, and RAMP requires no changes to existing backbone models for deployment.
Why It Matters
Helps advertisers maintain accurate targeting under privacy regulations like GDPR, balancing personalization with data protection requirements.