Async Learned User Embeddings for Ads Delivery Optimization

TL;DR

The paper tackles the challenge of delivering relevant ads by learning high-fidelity, time-aware user embeddings at scale. It introduces ALURE, an asynchronous pipeline that derives embeddings from multimodal sequence data via a Transformer-like module, compresses them for storage, and constructs user similarity graphs to enhance candidate retrieval. Offline evaluations show consistent gains (0.1–0.3% NE) when using embeddings as features, while online A/B tests demonstrate a 0.28% improvement in delivery performance with the ALURE-based graph. The approach enables scalable, timelier personalization in ads systems by integrating async learning, graph learning, and retrieval augmentation, potentially reducing real-time compute while boosting relevance.

Abstract

In recommendation systems, high-quality user embeddings can capture subtle preferences, enable precise similarity calculations, and adapt to changing preferences over time to maintain relevance. The effectiveness of recommendation systems depends on the quality of user embedding. We propose to asynchronously learn high fidelity user embeddings for billions of users each day from sequence based multimodal user activities through a Transformer-like large scale feature learning module. The async learned user representations embeddings (ALURE) are further converted to user similarity graphs through graph learning and then combined with user realtime activities to retrieval highly related ads candidates for the ads delivery system. Our method shows significant gains in both offline and online experiments.

Figures (8)

• Figure 1: Ads delivery funnel with multi-stage ranking
• Figure 2: The feedback loop for User Representation Learning
• Figure 3: An example of user sequence feature
• Figure 4: Illustration of user representation embedding learning architecture
• Figure 5: One layer of our custom Transformer-like architecture