Explaining Federated Learning-Based Movie Recommendations

Abstract

The widespread adoption of recommender systems across industries such as entertainment, healthcare, and e-commerce has heightened concerns related to privacy, transparency, and overall trustability of these systems. The traditional centralized recommender systems risk mismanagement of user privacy and they lack interpretability, conflicting with emerging regulatory standards outlined by the EU AI Act and EU Data Act. To address these arising challenges head on, we propose an innovative recommender system framework that blends concepts from Federated Learning (FL) for the aspects of privacy and Explainable AI (XAI) to increase system trustability through transparency. FL facilitates decentralized model training, preserving user privacy by ensuring that personal data remains on local user devices while aggregating the global model updates with data scrubbed of information centrally. To enhance transparency and user trust, we borrow the post-hoc explanation strategies from the XAI literature and we leverage Large Language Models (LLMs) to harmonize the explanations with clear, understandable sentences for recommendations toward end-users. This combined solution balances privacy preservation, regulatory compliance, personalized recommendations, and interpretability, significantly enhancing recommender system design and adoption.