Exact and Efficient Unlearning for Large Language Model-Based Recommendation

Abstract

Recent years have witnessed the trend of enhancing recommender systems with large language models (LLMs), namely, LLMRec. A common way is to fine-tune the LLMs with the instruction data transformed from user behaviors, stimulating the recommendation ability of LLMs. Similar to traditional recommender systems, integrating user data into LLMs raises privacy concerns. Users desire a tool to erase the impacts of their sensitive data from the trained models. To meet this user demand, LLMRec unlearning becomes pivotal to enable the removal of unusable data (e.g., historical behaviors) from established LLMRec models. However, existing methods mostly focus on partition strategies and approximate unlearning. These methods are not well-suited for the unique characteristics of LLMRec due to computational costs or incomplete removal. In this study, we propose the Adapter Partition and Aggregation (APA) framework for exact and efficient LLMRec unlearning while maintaining recommendation performance. APA achieves this by retraining PEFT adapters using data partitioning, constructing adapters for partitioned training data shards, and retraining only the affected adapters. To preserve recommendation performance and avoid significant inference costs, APA incorporates balanced and heterogeneous data partitioning, and parameter-level adapter aggregation with sample-adaptive adapter attention for each testing sample. Extensive experiments demonstrate the effectiveness and efficiency of our method.