Personalization in federated recommendation system using SVD++ with explainability

Abstract

Large-scale distributed Artificial Intelligence (AI) systems are getting more widespread as traditional AI applications require centralizing large amounts of data for training models, posing privacy and security risks. For this reason, the idea of Federated Learning (FL) has emerged where instead of sharing data, the edge devices send model parameters over the network to the global model. Though FL ensures privacy preservation, this system lacks personalization due to the heterogeneous data across the client devices. At the same time, the debate continues over the explainability of the FL model like other AI systems. This paper has implemented SVD++ for movie recommendations using the Movielens 10M dataset to increase personalization in the FL system. Later we have also inaugurated explainability to remove the black-box nature of the recommendation system. To our knowledge, implementing SDV++ for personalization in a federated learning setup has not been introduced before. Our trained model has achieved RMSE value of 0.8906. Finally, ensuring the principles of Responsible AI will make the FL recommendation system more fair and reliable.