.png){kind=icon}
.png){kind=icon}
Return
Active learning for LLM-based recommender systems
.png){kind=icon}
PRE
AI
DOI：10.1007/s10844-026-01048-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Amid the rapid advancement of Large Language Models (LLMs), recommender systems are undergoing a transformative shift. Traditional recommendation approaches struggle to deeply capture user intent and exhibit limited generalization ability in scenarios with sparse data. LLMs such as GPT-4 and LLaMA, known for their rich world knowledge and powerful reasoning capabilities, have attracted increasing attention from researchers exploring their applications in recommendation tasks. However, due to the significant differences between natural language processing and personalized recommendation, LLMs cannot be directly applied to recommendation scenarios. Although techniques like instruction tuning and parameter-efficient fine-tuning have alleviated this issue to some extent, they still face challenges such as low sample utilization efficiency and high resource consumption. To address these issues, this paper proposes an active learning framework for LLM-based recommender systems-ALLRec. This method evaluates the informativeness of training samples based on the loss values generated during the inference phase of LLMs, dynamically selecting key samples to improve training efficiency. Furthermore, we design a diversity learning module decoupled from the main LLMs, which captures structural and semantic differences between samples to enhance the diversity of selected samples in the data distribution. Under controllable resource constraints, ALLRec effectively identifies high-value samples by jointly modeling informativeness and diversity. Experimental results demonstrate that ALLRec significantly accelerates model convergence and improves recommendation performance across multiple benchmark datasets, validating its potential in LLM-based recommender systems.
Keywords:
Active learning
LLM-based recommender systems
Diversity
Training efficiency
Journal
J
IF:
3.4
Papers:
75
Citations:
0