LLM for (chat-based) Recommender Systems

Recommender systems often depend on collaborative filtering (CF) models, which face difficulties in effectively capturing semantic relationships while operating with profiles and adapting to changing user preferences. Large Language Models are well suited to solve emerging problem with text data to improve CF models; however, their need for frequent updates can be a barrier to achieving online scalability. The goal of this work was to investigate the possibility of transferring knowledge from LLM to an arbitrary CF model without the need for frequent calls to LLM during inference, resulting in the development of a new model-agnostic approach called “LLM-as-a-Guide”. The general idea is to train a CF model to recover features at a selected intermediate layer by embedding CF models into the encoder-decoder structure. The presented method is evaluated under two scenarios: online and offline scenarios. The results demonstrate significant improvements in both dynamic update performance and generalization across various CF models and datasets. This integration of LLM-generated features into CF models has led to substantial improvements in the overall performance and capabilities of the recommendation system while enabling more dynamic and adaptive updates to user preferences.