Retrieval Augmentation via User Interest Clustering

TL;DR

This work tackles data sparsity for light users and the fragmented interests of heavy users in industrial recommender systems by introducing an intermediate User Interest Clustering (UIC) layer. UIC builds an item co-engagement graph, clusters items into $K$ interests via Louvain, derives per-user interest vectors $\boldsymbol{\eta}_u$, and fuses these with user representations, enabling interest-level attention. Inference is performed on a restricted set of clusters using KNN, reducing complexity to $O\left( N \cdot \frac{|oldsymbol{\mathcal{I}}|}{K} \right)$ per user while maintaining or improving accuracy, and the approach is trained as a standard two-tower model with BCE loss. Empirical results on MovieLens-1M and Recipe show favorable accuracy–efficiency trade-offs, and production deployment at Meta demonstrates real-world improvements in short-form video recommendations with scalable inference and reduced latency.

Abstract

Many existing industrial recommender systems are sensitive to the patterns of user-item engagement. Light users, who interact less frequently, correspond to a data sparsity problem, making it difficult for the system to accurately learn and represent their preferences. On the other hand, heavy users with rich interaction history often demonstrate a variety of niche interests that are hard to be precisely captured under the standard "user-item" similarity measurement. Moreover, implementing these systems in an industrial environment necessitates that they are resource-efficient and scalable to process web-scale data under strict latency constraints. In this paper, we address these challenges by introducing an intermediate "interest" layer between users and items. We propose a novel approach that efficiently constructs user interest and facilitates low computational cost inference by clustering engagement graphs and incorporating user-interest attention. This method enhances the understanding of light users' preferences by linking them with heavy users. By integrating user-interest attention, our approach allows a more personalized similarity metric, adept at capturing the complex dynamics of user-item interactions. The use of interest as an intermediary layer fosters a balance between scalability and expressiveness in the model. Evaluations on two public datasets reveal that our method not only achieves improved recommendation performance but also demonstrates enhanced computational efficiency compared to item-level attention models. Our approach has also been deployed in multiple products at Meta, facilitating short-form video related recommendation.

Figures (5)

• Figure 1: (a) The top 10 most popular items account for almost half of all user-item interactions. (b) The basic two-tower model tends to recommend more popular items, even though users often interact with less popular ones.
• Figure 2: Recommendation quality varies significantly across users with different levels of engagement.
• Figure 3: UIC effectively addresses the challenges associated with serving both light and heavy users. The clustering approach based on engagement signals outperforms the one that relies solely on item features (i.e., Kmeans). The x-axis indicates the percentage of users with lower engagement levels than the referenced user.
• Figure 4: Leveraging KNN acceleration techniques reduces inference time significantly.
• Figure 5: The t-SNE visualization of the item vectors recommended to a sampled user in MovieLens-1M via UIC.