Exact and Efficient Unlearning for Large Language Model-based Recommendation

TL;DR

Privacy concerns in LLM-based recommendation arise from the need to remove data a model should forget. The paper introduces Adapter Partition and Aggregation (APA), an unlearning framework that partitions training data into shards, assigns distinct adapters to each shard, and retrains only the affected adapters to achieve exact data removal while maintaining performance. To keep inference costs low, APA employs parameter-level adapter aggregation with sample-adaptive attention during testing, enabling efficient, personalized predictions. Extensive experiments verify APA's ability to achieve exact unlearning with minimal impact on recommendation quality and improved efficiency, underscoring its practicality for privacy-preserving LLM-based recommendation.

Abstract

The evolving paradigm of Large Language Model-based Recommendation (LLMRec) customizes Large Language Models (LLMs) through parameter-efficient fine-tuning (PEFT) using recommendation data. The inclusion of user data in LLMs raises privacy concerns. To protect users, the unlearning process in LLMRec, specifically removing unusable data (e.g., historical behaviors) from established LLMRec models, becomes crucial. However, existing unlearning methods are insufficient for the unique characteristics of LLM-Rec, mainly due to high computational costs or incomplete data erasure. In this study, we introduce the Adapter Partition and Aggregation (APA) framework for exact and efficient unlearning while maintaining recommendation performance. APA achieves this by establishing distinct adapters for partitioned training data shards and retraining only the adapters impacted by unusable data for unlearning. To preserve recommendation performance and mitigate considerable inference costs, APA employs parameter-level adapter aggregation with sample-adaptive attention for individual testing samples. Extensive experiments substantiate the effectiveness and efficiency of our proposed framework