Machine Unlearning In Recommendation Systems

Machine unlearning is crucial for ensuring privacy and security in recommendation systems, allowing selective removal of user data without full model retraining. This paper explores exact and approximate unlearning techniques, particularly in graph-based recommendation models, where user-item interactions create challenges in data removal. A balanced clustering-based unlearning framework is introduced, partitioning user data into shards to optimize retraining efficiency. Experimental results show that this approach reduces computational costs by 85% while maintaining model accuracy within 1-2% of a fully retrained system. The proposed method ensures scalability, efficiency, and compliance with privacy regulations, making it well-suited for large-scale applications.