KNOWLIGHT: WHERE LARGE LANGUAGE MODELS MEET GRAPH RECOMMENDATION VIA KNOWLEDGE DISTILLATION

This thesis introduces KnowLight, a unified Knowledge Distillation (KD) framework designed to transfer the rich semantic knowledge of Large Language Models (LLMs) to the efficient Graph Neural Networks (GNNs) for modern recommendation systems. Its name encapsulates its core mission: to blend the semantic richness and reasoning capabilities of language models ("Know") with the extreme efficiency and topological scalability of LightGCN ("Light"). Traditional graph-based collaborative models, such as LightGCN, excel at capturing high-order co-occurrence patterns but lack textual understanding and suffer from performance degradation in high-sparsity or cold-start scenarios. Conversely, while LLMs have demonstrated remarkable capabilities in understanding metadata and reasoning about preferences, their computational complexity and inference latency preclude their direct deployment in high-throughput industrial recommender systems. KnowLight addresses this challenge by transferring latent knowledge from a Teacher LLM to a compact Student model (LightGCN), exploring different distillation channels. In doing so, KnowLight enhances the GNN’s discriminative power while preserving its full efficiency during inference.