Reproduce “Preliminary Study on Incremental Learning for Large Language Model-based Recommender Systems” on OpenResearch

About this paper

Tianhao Shi, Yang Zhang, Zhijian Xu, Chong Chen, Fuli Feng, Xiangnan He, Qi Tian

Researchers investigated incremental learning for Large Language Model-based Recommender Systems (LLM4Rec), finding that traditional methods did not improve performance. They introduced the Long- and Short-term Adaptation-aware Tuning (LSAT) framework, which leverages separate LoRA modules for stable long-term and dynamic short-term preferences, demonstrating improved performance over conventional incremental updates on two public datasets.

Key insights

• Traditional incremental learning methods (full retraining, fine-tuning) did not lead to performance improvements for LoRA-based LLM4Rec, unlike for traditional recommender models.
• LLM4Rec models exhibit strong generalization capabilities to new data, including cold-start items, which somewhat mitigates the immediate need for frequent updates seen in traditional systems.
• A single LoRA module struggles to effectively capture both long-term and short-term dynamic user preferences simultaneously, suggesting a capacity limitation in a dynamic incremental learning setting.

Summary by alphaXiv.