TechPrufer
Sign inGet started

Blogs›Uber›Ads Personalization with Sequential Modeling and Hetero-MMoE

Ads Personalization with Sequential Modeling and Hetero-MMoE

Ads Personalization with Sequential Modeling and Hetero-MMoE

1
posts
2026

Uber has enhanced its ads personalization system by incorporating sequential user behavior modeling with a target-aware transformer encoder and a novel Hetero-MMoE architecture. This evolution moves beyond static, aggregate features to capture temporal dynamics and evolving user intent using multi-head latent attention for efficient sequence summarization. The Hetero-MMoE framework blends MLP, DCN, and CIN experts to capture low- to high-order feature interactions more effectively, enabling more precise, context-aware ad predictions across diverse objectives like CTR and CTO. This architecture serves as a general foundation for multi-objective, multi-task modeling across the platform, leading to improved ad relevance, user engagement, and advertiser outcomes.

2026

Transforming Ads Personalization with Sequential Modeling and Hetero-MMoE at Uber

3/10/2026

This post details the technical advancements in Uber's ads personalization system, specifically the introduction of sequential user features using a target-aware transformer encoder with Multi-Head Latent Attention (MLA) for efficient sequence summarization, and the adoption of a Hetero-MMoE framework. MLA reduces attention complexity from O(N^2) to O(N*L) by using a fixed-size set of learnable latent tokens. The Hetero-MMoE integrates MLP, DCN, and CIN experts to capture diverse feature interactions, improving model capacity for multi-objective learning. The post also presents performance gains in pCTR and pCTO metrics.