Large language models make sample-efficient recommender systems

In the rapidly evolving digital era, the quest for personalized user experiences has led to an explosion in the development of recommender systems (RSs). These systems are the driving force behind the tailored content we encounter daily, from e-commercial product suggestions to personalized news feeds. However, the traditional models have often struggled with the challenge of sparsity and the need for vast amounts of data to achieve satisfactory performance. This is where the innovative research presented in “Large Language Models Make Sample-Efficient Recommender Systems” comes to the forefront, offering a groundbreaking approach to enhance the efficiency and effectiveness of RSs.

This paper is published on 15 Apr 2025  in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

Authored by a collaborative team of researchers from Shanghai Jiao Tong University and Huawei Noah’s Ark Lab, the paper introduces a novel framework named Laser, which integrates the prowess of LLMs into RSs. The authors argue that by leveraging the remarkable capabilities of LLMs in natural language processing (NLP), RSs can become significantly more sample-efficient, requiring a fraction of the data traditionally needed for training.

The core innovation lies in two key aspects: First, LLMs themselves can act as efficient recommenders, capable of understanding and predicting user preferences with limited data. Second, when used in conjunction with conventional recommendation models (CRMs), LLMs enhance their sample efficiency by serving as feature generators and encoders, thus improving overall performance with a smaller dataset.

The research demonstrates that the Laser framework, through extensive experiments on public datasets, can match or even surpass the performance of CRMs trained on complete datasets, using only a small fraction of the training samples. The paper also addresses potential concerns regarding the practicality of deploying LLMs in real-world applications. The authors discuss the inference latency challenges and propose solutions to ensure that the integration of LLMs into RSs meets the strict latency requirements of industrial applications. The Laser framework’s design allows for the pre-storage of LLM-augmented user and item representations, effectively mitigating the latency issues associated with real-time LLM processing.

The authors conclude by highlighting the potential for further improvements in sample efficiency, suggesting strategies such as selective sampling for training and exploring the application of Laser in other domains like code snippet recommendation, which offers a path towards more intelligent, efficient, and personalized recommendation systems.

DOI: 10.1007/s11704-024-40039-z