In a groundbreaking advancement for privacy and utility in recommendation systems, researchers have developed a novel method for unlearning in Large Language Models (LLMs) used for recommendations. This pioneering approach, known as E2URec, addresses the critical need for LLM-based recommenders to forget specific user data while maintaining recommendation performance, a challenge that has been largely unaddressed in the field. This is particularly significant in the era of data-sensitive applications where the ‘Right To Be Forgotten’ is increasingly recognized. Their study, which could revolutionize how recommendation systems handle privacy requests and noisy data, was published on 15 Mar 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
Led by a collaborative team from Shanghai Jiao Tong University and Huawei Noah’s Ark Lab, the research introduces a lightweight, low-rank adaptation module that updates only a fraction of the model's parameters, significantly reducing the computational cost and time associated with unlearning. The E2URec method also employs a pair of teacher models to guide the unlearned model in forgetting irrelevant data without compromising its recommendation capabilities.
The originality of this research lies in its dual approach to unlearning, which not only ensures that the model forgets the specified data but also retains its overall recommendation performance. This is achieved through the innovative design of a forgetting teacher that approximates the output of a retrained model devoid of the forgotten data, and a remembering teacher that ensures the model's output on retained data remains consistent with the original, high-performing model.
The team conducted comprehensive experiments on public recommendation datasets, demonstrating that E2URec outperforms existing methods in both efficiency and effectiveness. The results indicate that E2URec maintains high recommendation performance metrics such as AUC and Accuracy, while drastically reducing the time and parameters required for unlearning.
The implications of this study extend beyond the technical realm, offering a solution that aligns with increasing privacy regulations and user expectations for data control. The E2URec method sets a new standard for the responsible and efficient management of data in LLM-based recommendation systems.
DOI： 10.1007/s11704-024-40044-2