Privacy-Preserving Feature Attribution Explanations for Large-Scale Recommendation Systems: A Differential Privacy Approach

Abstract

Modern recommendation systems increasingly demand explainable predictions while simultaneously protecting user privacy. Existing feature attribution methods for recommender systems often expose sensitive user information through detailed explanations, creating significant privacy risks. This paper presents a comprehensive privacy-preserving feature attribution framework specifically designed for large-scale recommendation systems. Our approach integrates differential privacy mechanisms with gradient-based feature attribution techniques, enabling transparent recommendations while maintaining strict privacy guarantees. The framework employs adaptive noise injection, dynamic privacy budget allocation, and multi-level transparency controls to balance explanation quality with privacy protection. We introduce novel concentrated differential privacy composition bounds optimized for sequential attribution queries and auto-mated compliance verification mechanisms. Extensive experiments on MovieLens, Amazon, and Yelp datasets demonstrate that our framework maintains reasonable recommendation accuracy while providing meaningful explanations under strong privacy constraints. The proposed approach achieves privacy-utility trade-offs with recommendation accuracy degradation of 8-15% while ensuring ε-differential privacy with ε ≤ 1.0, representing a significant improvement over existing privacy-preserving explanation method.