Synergizing Implicit and Explicit User Interests: A Multi-Embedding Retrieval Framework at Pinterest

TL;DR

This paper introduces a multi-embedding retrieval framework that combines implicit and explicit user interests to improve candidate retrieval in Pinterest's recommendation system, leading to higher engagement and diversity.

Contribution

The paper presents a novel multi-embedding retrieval framework that effectively integrates implicit and explicit user interests for enhanced recommendation retrieval.

Findings

Significant improvements in user engagement metrics.

Enhanced feed diversity observed in experiments.

Successful deployment on Pinterest's home feed.

Abstract

Industrial recommendation systems are typically composed of multiple stages, including retrieval, ranking, and blending. The retrieval stage plays a critical role in generating a high-recall set of candidate items that covers a wide range of diverse user interests. Effectively covering the diverse and long-tail user interests within this stage poses a significant challenge: traditional two-tower models struggle in this regard due to limited user-item feature interaction and often bias towards top use cases. To address these issues, we propose a novel multi-embedding retrieval framework designed to enhance user interest representation by generating multiple user embeddings conditioned on both implicit and explicit user interests. Implicit interests are captured from user history through a Differentiable Clustering Module (DCM), whereas explicit interests, such as topics that the user has followed, are modeled via Conditional Retrieval (CR). These methodologies represent a form of conditioned user representation learning that involves condition representation construction and associating the target item with the relevant conditions. Synergizing implicit and explicit user interests serves as a complementary approach to achieve more effective and comprehensive candidate retrieval as they benefit on different user segments and extract conditions from different but supplementary sources. Extensive experiments and A/B testing reveal significant improvements in user engagements and feed diversity metrics. Our proposed framework has been successfully deployed on Pinterest home feed.