kNN-Embed: Locally Smoothed Embedding Mixtures For Multi-interest Candidate Retrieval

TL;DR

kNN-Embed enhances candidate retrieval in recommendation systems by modeling user interests as a smoothed mixture over learned clusters, significantly improving diversity and recall over standard methods.

Contribution

This paper introduces kNN-Embed, a novel method that models user interests as a mixture over learned clusters to improve diversity in dense ANN-based candidate retrieval.

Findings

Significant improvements in recall across three datasets.

Enhanced diversity in retrieved candidate sets.

Open-sourced a large Twitter follow-graph dataset.

Abstract

Candidate retrieval is the first stage in recommendation systems, where a light-weight system is used to retrieve potentially relevant items for an input user. These candidate items are then ranked and pruned in later stages of recommender systems using a more complex ranking model. As the top of the recommendation funnel, it is important to retrieve a high-recall candidate set to feed into downstream ranking models. A common approach is to leverage approximate nearest neighbor (ANN) search from a single dense query embedding; however, this approach this can yield a low-diversity result set with many near duplicates. As users often have multiple interests, candidate retrieval should ideally return a diverse set of candidates reflective of the user's multiple interests. To this end, we introduce kNN-Embed, a general approach to improving diversity in dense ANN-based retrieval. kNN-Embed represents each user as a smoothed mixture over learned item clusters that represent distinct "interests" of the user. By querying each of a user's mixture component in proportion to their mixture weights, we retrieve a high-diversity set of candidates reflecting elements from each of a user's interests. We experimentally compare kNN-Embed to standard ANN candidate retrieval, and show significant improvements in overall recall and improved diversity across three datasets. Accompanying this work, we open source a large Twitter follow-graph dataset (https://huggingface.co/datasets/Twitter/TwitterFollowGraph), to spur further research in graph-mining and representation learning for recommender systems.