LLMs for User Interest Exploration: A Hybrid Approach

Jianling Wang, Haokai Lu, Yifan Liu, He Ma, Yueqi Wang, Yang Gu, Shuzhou Zhang, Ningren Han, Shuchao Bi, Lexi Baugher, Ed H. Chi, Minmin Chen

Abstract

Researchers at Google DeepMind have developed a hybrid hierarchical framework that leverages Large Language Models (LLMs) to overcome the "feedback loop" inherent in large-scale recommendation systems. By utilizing a fine-tuned LLM to propose high-level "novel interest clusters" and grounding these suggestions in traditional item-level models, the system successfully increases both user interest exploration and overall platform engagement for billions of users.

Introduction

Traditional recommendation systems often fall victim to an "echo chamber" effect. By learning primarily from past user-item interactions, these systems create a strong feedback loop: they recommend what a user already likes, the user clicks, and the system reinforces that specific preference. While effective for short-term engagement, this limits the discovery of novel interests and can lead to content fatigue.

Large Language Models (LLMs) possess vast world knowledge that could theoretically suggest new topics. However, deploying them in industrial-scale environments (e.g., YouTube, Spotify) presents significant challenges: LLMs are often too slow for sub-100ms response times, they lack knowledge of specific, rapidly changing private catalog IDs, and they may fail to capture the nuanced collaborative signals that classic models handle with ease.

Methodology: Hybrid Hierarchical Planning

The proposed solution is a hybrid hierarchical planning approach that decouples the recommendation task into two distinct levels:

  1. High-level Language Policy (The "What"): Instead of predicting specific items, an LLM (Gemini-Pro) predicts the next interest cluster. Billions of items are grouped into topical clusters (e.g., "Underwater diving" or "Bus simulators"). The LLM processes a user's recent history as a sequence of cluster descriptions and generates a description for a novel cluster the user might enjoy.
  2. Low-level Item Policy (The "Which"): Once a cluster is selected, a classic transformer-based sequential model takes over. Crucially, its output is restricted to items within the LLM-selected cluster, ensuring the recommendation is both novel (via the LLM) and personalized (via the classic model).
The hybrid hierarchical framework combining high-level language policies with low-level item policies.

Implementation Details

To ensure the LLM generates valid cluster names, the team used Supervised Fine-Tuning (SFT). This ensures the model outputs exact matches for the system's internal database (e.g., "Underwater diving" instead of "Scuba diving"). To solve the latency bottleneck, the system limits user history context to the last two clusters ($K=2$), allowing for the pre-computation of cluster-to-cluster transitions in an offline lookup table.

Experimental Setup and Training

The study was conducted on a commercial short-form video platform serving billions of users. The SFT training data consisted of 250,000 successful "novel interest transitions" from real user logs. To prevent the LLM from merely acting as a popularity recommender, the researchers implemented label balancing, ensuring the model could navigate the "long tail" of niche interests by providing equal representation for all 761 possible interest clusters in the training set.

Key Results

Live A/B experiments demonstrated that the hybrid model significantly outperforms state-of-the-art exploration models like contextual bandits.

  • Technical Accuracy: The fine-tuned LLM achieved a match rate exceeding 99%, confirming the effectiveness of the SFT approach.
  • User Interest Expansion: The system showed a marked increase in UCI@N (Unique Clusters of Interest), indicating that users were exploring a wider variety of topics.
  • Platform Engagement: The framework drove measurable increases in overall watch time and the number of active users (specifically those with >10 minutes of daily watch time), proving that novelty, when properly grounded, enhances long-term satisfaction.
Comparison of the LLM-powered approach against exploration-oriented and exploitation-oriented baselines in terms of novelty and positive feedback.

Conclusion

This research demonstrates that LLMs are most effective in recommendation systems when used as "creative planners" rather than direct item predictors. By using natural language as the interface between high-level intent and low-level execution, the framework leverages the world knowledge of LLMs while maintaining the speed and precision of classic models. Future work will explore expanding the context window beyond $K=2$ and assessing the long-term lifetime value of these novel interest introductions.

Summary updated

BibTeX
@inproceedings{wang2024llms,
  author = {Jianling Wang and Haokai Lu and Yifan Liu and He Ma and Yueqi Wang and Yang Gu and Shuzhou Zhang and Ningren Han and Shuchao Bi and Lexi Baugher and Ed H. Chi and Minmin Chen},
  title = {{LLMs for User Interest Exploration: A Hybrid Approach}},
  booktitle = {RecSys},
  year = {2024},
  doi = {10.1145/3640457.3688161},
}
Figure