Integrating Prior Knowledge from Meta-Learning and Large Language Models for Cold-Start Recommendation

Abstract

Recommender Systems (RSs) often suffer from the cold-start problem, which leads to poor recommendations when dealing with new items or users with limited interaction histories. Recent research has leveraged meta-learning to overcome this issue by extracting informative prior knowledge from the interaction histories of warm users. This prior knowledge helps recommenders identify the preferences of new users with only a few interactions. In this work, we tackle the cold-start problem using two types of prior knowledge: recommendation-specific knowledge derived from meta-learning and general knowledge obtained from Large Language Models (LLMs) pre-trained on vast amounts of internet data. Specifically, our approach involves encoding item and user textual information using LLMs, allowing us to effectively incorporate general prior knowledge. Additionally, we train the recommender using meta-learning techniques to acquire recommendation-specific prior knowledge. By combining these two types of knowledge, our method demonstrates promising performance under various cold-start settings.