COFFEE: COdesign Framework for Feature Enriched Embeddings in Ads-Ranking Systems

TL;DR

COFFEE tackles the challenge of building richer user–ad representations from diverse, multi-modal data without increasing inference cost. It introduces a three‑dimensional co-design framework for feature enriched embeddings (EBF Sequences) that expands in the number of events, sequence length, and semantic content, implemented via a modified Ads‑Recommendation Model with an event module. The study finds that ad‑impression EBFs with enrichment yield approximately $1.56\times$ higher AUC and $1.52\times$ steeper training‑capacity slopes, with a $0.56\%$ absolute CTR AUC gain at modest online sequence lengths, and scalability to offline sequences on the order of $10^6$ events. Qualitative analyses show cross‑attention can explain the relevance of EBF signals to served ads, supporting personalized EBFs beyond historic engagement. The results suggest practical gains in CTR prediction and scalable personalization using multi‑modal embeddings while maintaining serving efficiency, and point to future work with compressed LLM embeddings for longer histories.

Abstract

Diverse and enriched data sources are essential for commercial ads-recommendation models to accurately assess user interest both before and after engagement with content. While extended user-engagement histories can improve the prediction of user interests, it is equally important to embed activity sequences from multiple sources to ensure freshness of user and ad-representations, following scaling law principles. In this paper, we present a novel three-dimensional framework for enhancing user-ad representations without increasing model inference or serving complexity. The first dimension examines the impact of incorporating diverse event sources, the second considers the benefits of longer user histories, and the third focuses on enriching data with additional event attributes and multi-modal embeddings. We assess the return on investment (ROI) of our source enrichment framework by comparing organic user engagement sources, such as content viewing, with ad-impression sources. The proposed method can boost the area under curve (AUC) and the slope of scaling curves for ad-impression sources by 1.56 to 2 times compared to organic usage sources even for short online-sequence lengths of 100 to 10K. Additionally, click-through rate (CTR) prediction improves by 0.56% AUC over the baseline production ad-recommendation system when using enriched ad-impression event sources, leading to improved sequence scaling resolutions for longer and offline user-ad representations.

Figures (5)

• Figure 1: Proposed 3-dimensional framework to enrich online-EBF sequences with high adaptability to user-intent.
• Figure 2: Representations for the Sparse NN-based models (left) vs. the Sequence learning Recommendation model (right).
• Figure 3: Training capacity scaling curves for EBF sources with varying sequence lengths and enrichment-levels. Ad-impression is the strongest source (in green and orange), followed by organic-impression (blue) and video-view (red). The baseline production model performance is depicted by the origin data point in black.
• Figure 4: Improved ad-prediction using organic impression EBF source vs. historic baseline signals.
• Figure 5: User-intent identification for new-ads using organic sequences over baseline.