Preliminary Study on Incremental Learning for Large Language Model-based Recommender Systems

TL;DR

The paper investigates incremental learning for Large Language Model-based recommender systems ($LLM4Rec$) and finds that standard strategies like full retraining and fine-tuning provide limited gains for LoRA-based configurations. To reconcile the mismatch between LLM4Rec architectures and incremental updates, it introduces LSAT, a two-adapter framework with separate long-term and short-term LoRA modules, coupled with two inference strategies (ensemble and task arithmetic). Empirical results on MovieLens-1M and Amazon-Book show that LSAT variants consistently outperform full retraining and fine-tuning, validating the long-term vs short-term adaptation hypothesis. The work highlights practical implications for real-world deployment of $LLM4Rec$ and provides code for replication, while also noting that future work should explore broader architectural directions beyond the current LoRA-centric view.

Abstract

Adapting Large Language Models for Recommendation (LLM4Rec) has shown promising results. However, the challenges of deploying LLM4Rec in real-world scenarios remain largely unexplored. In particular, recommender models need incremental adaptation to evolving user preferences, while the suitability of traditional incremental learning methods within LLM4Rec remains ambiguous due to the unique characteristics of Large Language Models (LLMs). In this study, we empirically evaluate two commonly employed incremental learning strategies (full retraining and fine-tuning) for LLM4Rec. Surprisingly, neither approach shows significant improvements in the performance of LLM4Rec. Instead of dismissing the role of incremental learning, we attribute the lack of anticipated performance enhancement to a mismatch between the LLM4Rec architecture and incremental learning: LLM4Rec employs a single adaptation module for learning recommendations, limiting its ability to simultaneously capture long-term and short-term user preferences in the incremental learning context. To test this speculation, we introduce a Long- and Short-term Adaptation-aware Tuning (LSAT) framework for incremental learning in LLM4Rec. Unlike the single adaptation module approach, LSAT utilizes two distinct adaptation modules to independently learn long-term and short-term user preferences. Empirical results verify that LSAT enhances performance, thereby validating our speculation. We release our code at: https://github.com/TianhaoShi2001/LSAT.

Figures (5)

• Figure 1: Incremental learning process in recommendation.
• Figure 2: Performance of TALLRec, BookGPT and traditional models obtained at different update periods on $\mathcal{D}_{20}$.
• Figure 3: Performance comparison between TALLRec and baselines on warm items and cold items. All models are trained on $\mathcal{D}_{1}-\mathcal{D}_{15}$ and tested on $\mathcal{D}_{16}-\mathcal{D}_{20}$.
• Figure 4: Performance of TALLRec and traditional models obtained at different update periods for warm items on $\mathcal{D}_{20}$.
• Figure 5: Performance comparison of full retraining, fine-tuning, and LSAT. All models are updated promptly with newly collected data $\mathcal{D}_{t}$ and tested on $\mathcal{D}_{t+1}$. LSAT (10) means it utilizes $\mathcal{H} =\left \{\mathcal{D}_1, \mathcal{D}_2, \dots, \mathcal{D}_{10} \right \}$ to train the long-term LoRA.