LLMs for User Interest Exploration in Large-scale Recommendation Systems

TL;DR

Traditional recommender systems suffer from feedback loops that limit discovery of novel interests. The authors propose a hybrid hierarchical planning framework that uses LLMs to generate novel interests within predefined topic clusters and grounds them to items via a cluster-constrained transformer-based recommender, complemented by supervised fine-tuning and offline precomputation to meet latency constraints. Key contributions include a cluster-based language policy for controlled generation within a fixed taxonomy, diversified fine-tuning data to align with actual user transitions and reduce long-tail effects, and offline precomputation of cluster-pair transitions ($M_2^2$) enabling fast online lookup, validated by live deployment on a platform with billions of users that shows increased novelty and user engagement. Overall, the work offers a scalable, production-ready method to broaden user interests while preserving personalization in industrial-scale systems.

Abstract

Traditional recommendation systems are subject to a strong feedback loop by learning from and reinforcing past user-item interactions, which in turn limits the discovery of novel user interests. To address this, we introduce a hybrid hierarchical framework combining Large Language Models (LLMs) and classic recommendation models for user interest exploration. The framework controls the interfacing between the LLMs and the classic recommendation models through "interest clusters", the granularity of which can be explicitly determined by algorithm designers. It recommends the next novel interests by first representing "interest clusters" using language, and employs a fine-tuned LLM to generate novel interest descriptions that are strictly within these predefined clusters. At the low level, it grounds these generated interests to an item-level policy by restricting classic recommendation models, in this case a transformer-based sequence recommender to return items that fall within the novel clusters generated at the high level. We showcase the efficacy of this approach on an industrial-scale commercial platform serving billions of users. Live experiments show a significant increase in both exploration of novel interests and overall user enjoyment of the platform.

Figures (5)

• Figure 1: LLM-powered hybrid hierarchical planning diagram for user interest exploration.
• Figure 2: Prompt for Novel Interest Prediction when $K=2$.
• Figure 3: Label (i.e., generated by fine-tuned LLM) Distribution: X-axis represents label frequency; Y-axis represents the percentage of labels within each frequency range.
• Figure 4: (a) Model Finetuning Process. (b) and (c) Comparison between different recommenders in live experiments.
• Figure 5: The proposed method drives user growth.