LCG-Rec is a novel two-stage recommendation framework that leverages Large Language Models (LLMs) as candidate generators rather than traditional retrievers. In few-shot settings, conventional candidate pools often lack diversity and fail to surface hard negatives—both essential for effective ranker training. LCG-Rec addresses this by generating virtual candidates in natural language, tailored to the user's preference and history, without relying on a fixed item corpus.
Our pipeline includes:
- LLM-based generation of a large candidate pool conditioned on user preferences
- MMR filtering to ensure semantic diversity
- Distance Binning, which balances candidate difficulty (Easy / Medium / Hard) based on embedding distance
This process results in hard and diverse candidates, improving both generalization and training stability.
We evaluate LCG-Rec on four POI benchmarks under 5% few-shot splits. The method significantly outperforms retrieval-based baselines, achieving +80% relative improvement in HR@1 on Yelp2018.
