UniPinRec: Unifying Generative Retrieval and Ranking at Pinterest Scale

2026-05-29

AI Summary

This work addresses the computational, parametric, and deployment redundancies arising from the conventional separation of retrieval and ranking modules in recommender systems. It proposes the first production-grade, end-to-end unified framework that integrates input processing, modeling, training, and deployment into a single stack. Built upon a shared Transformer backbone, the framework introduces three key techniques: Masked Action Modeling (MAM), hybrid training sample construction, and cross-stage key-value cache reuse, which jointly enable approximate nearest neighbor retrieval and cross-attention-based ranking within a cohesive architecture. Online experiments demonstrate that the proposed method improves user engagement by approximately 1%, reduces end-to-end latency by 11.1%, and increases system throughput (QPS) by 63.6%.

Abstract

Modern recommendation systems predominantly train retrieval and ranking as separate models despite both increasingly relying on large transformers encoding the same user behavior data, duplicating parameters, compute, and serving cost. Prior work unifies the model architecture but not the full pipeline: input formats, training procedures, and serving stacks remain fragmented across stages. We present UniPinRec, which achieves full-stack unification of retrieval and ranking at Pinterest: one input format, one model, one training stage, deployed within existing serving infrastructure. A shared transformer encodes the user action sequence into candidate-independent representations that branch into retrieval (ANN dot-product) and ranking (cross-attention) via task-specific heads. Three ideas make this work: (1) Masked Action Modeling (MAM) eliminates interleaving, enabling weight sharing without doubling context length; (2) Blended training examples pair action sequences with feedview impression slates to satisfy both objectives jointly; (3) Cross-stage KV cache sharing reuses user-history computation from retrieval for ranking, reducing total FLOPs versus serving two independent models. Deployed in the Pinterest core surfaces, UniPinRec delivers approximately +1% online engagement lift while cutting end-to-end serving latency by 11.1% and lifting QPS by 63.6%. To our knowledge, this is the first full-stack unification of retrieval and ranking, covering inputs, model, training and serving, deployed in a production recommendation system.

Problem

• retrieval
• ranking
• recommendation systems
• model unification
• serving efficiency

Innovation

• Unified Retrieval and Ranking
• Masked Action Modeling
• KV Cache Sharing
• Transformer-based Recommendation
• End-to-End Recommendation System