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.

Recommender systems are essential for finding personalized content for users on online platforms. These systems are often trained on historical user interaction data, which collects user feedback on system recommendations. This creates a feedback loop leading to popularity bias; popular content is over-represented in the data, better learned, and thus recommended even more. Less popular content struggles to reach its potential audiences. Popularity bias limits the diversity of content that users are exposed to, and makes it harder for new creators to gain traction. Existing methods to alleviate popularity bias tend to trade off the performance of popular items. In this work, we propose a new method for alleviating popularity bias in recommender systems, called the cluster anchor regularization, which partitions the large item corpus into hierarchical clusters, and then leverages the cluster information of each item to facilitate transfer learning from head items to tail items. Our results demonstrate the effectiveness of the proposed method with offline analyses and live experiments on a large-scale industrial recommendation platform, where it significantly increases tail recommendation without hurting the overall user experience.