Enhancing Ad Relevance: Integrating Real-Time Context into Sequential Recommender Models

2026-05-08 · Source: Pinterest Engineering Blog - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Recommender Systems · Depth: Advanced, medium

Summary

Pinterest engineers introduced the Contextual Sequential Two Tower Model on May 8, 2026, significantly enhancing ad relevance by integrating real-time online context into their Transformer-based sequential recommender. The prior model, relying only on offline user conversion history, struggled on contextual surfaces like Related Pins, attributing less than 1% of impressions. The new architecture incorporates a context layer into the query tower, combining real-time subject Pin features, such as aggregated interest categories, with historical sequence data. Training utilizes synthetic augmented data from positive labels, while inference employs a hybrid approach, merging offline Transformer encoding with online context layer computation. This model achieved a 3x to 10x increase in Recall@K during offline evaluation. Online, it boosted median candidate relevance by ~275–300%, improved overall Related Pins ad relevance by 1.08%, and delivered a ~0.7% lift in Return on Ad Spend (ROAS), reaching ~1.4% in top countries.

Key takeaway

For AI Engineers optimizing ad recommendation systems on contextual surfaces, you should prioritize integrating real-time user context. Your existing sequential models, if purely offline, likely underperform on surfaces like Related Pins. By adopting a hybrid inference approach and training with synthetic augmented data, you can achieve significant lifts in ad relevance and Return on Ad Spend. Explore advanced fusion techniques like cross-attention to further enhance contextual signal integration.

Key insights

Integrating real-time context into sequential recommenders significantly boosts ad relevance and business performance.

Principles

• Offline historical data alone limits contextual relevance.
• Synthetic data enables training on online-only features.
• Hybrid inference combines pre-computed and real-time signals.

Method

A Contextual Sequential Two Tower Model integrates a context layer into the query tower, concatenating real-time features with Transformer output, trained via synthetic augmented data, and uses hybrid offline/online inference.

In practice

• Augment user embeddings with demographic features.
• Apply high dropout in the context layer during training.
• Investigate cross-attention for context fusion.

Topics

• Ad Relevance
• Sequential Recommender Models
• Real-Time Context
• Two-Tower Models
• Hybrid Inference
• Synthetic Data Augmentation

Best for: Machine Learning Engineer, AI Engineer, AI Architect

Related on AIssential

• Google Search’s AI evolution includes more ads — Google is integrating AI into search ads to create personalized, conversational, and contextually relevant product promotions.
• In-Context Multiple Instance Learning — Pretraining an in-context learner on synthetic data solves low-label MIL tasks efficiently without gradient updates.
• Ads are coming to AI. Does that really have to be such a bad thing? — Transparent, contextual advertising in AI can fund accessibility and enhance user experience without compromising trust.
• Milliseconds to Match: Criteo's AdTech AI & the Future of Commerce w/ Diarmuid Gill & Liva Ralaivola — Modern ad tech relies on millisecond-speed AI, deep learning, and foundation models for real-time, privacy-conscious recommendations.
• TokenFormer: Unify the Multi-Field and Sequential Recommendation Worlds — TokenFormer unifies multi-field and sequential recommendation by preventing "Sequential Collapse Propagation" through novel attention and representation schemes.
• X-SYNTH: Beyond Retrieval -- Enterprise Context Synthesis from Observed Human Attention — Human attention, captured as ordered digital interaction sequences, provides a reliable relevance signal for enterprise AI agents.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Pinterest Engineering Blog - Medium.