TRAWL: External Knowledge-Enhanced Recommendation with LLM Assistance
Weiqing Luo, Chonggang Song, Lingling Yi, Gong Cheng
TL;DR
TRAWL tackles the challenge of augmenting recommender systems with external knowledge by using large language models to extract recommendation-focused knowledge from unstructured sources and by learning a semantic-to-behavioral adapter through contrastive, multi-task training. The framework splits into a knowledge-generation module that produces user- and item-specific semantic content guided by key factors, and a knowledge-adaptation module that freezes the semantic encoder while training an adapter to inject behavioral signals into the recommendation space. Empirical results on MovieLens-1M across several backbones show clear gains over baselines, and an industrial WeChat deployment demonstrates real-world improvements in clicks, engagement time, and interactions. Overall, TRAWL provides a versatile, scalable approach to fuse semantic external knowledge with behavioral data in recommender systems, with strong potential for broader deployment and further enhancement via prompt design and adaptation strategies.
Abstract
Combining semantic information with behavioral data is a crucial research area in recommender systems. A promising approach involves leveraging external knowledge to enrich behavioral-based recommender systems with abundant semantic information. However, this approach faces two primary challenges: denoising raw external knowledge and adapting semantic representations. To address these challenges, we propose an External Knowledge-Enhanced Recommendation method with LLM Assistance (TRAWL). This method utilizes large language models (LLMs) to extract relevant recommendation knowledge from raw external data and employs a contrastive learning strategy for adapter training. Experiments on public datasets and real-world online recommender systems validate the effectiveness of our approach.