Revolutionizing Personal Recommendations via Federated Contrastive Transformer Learning

This paper presents an innovative solution for personalized recommendations in consumer electronics, addressing data confidentiality through federated learning mechanism, and tackling data quantity and quality using transformers and consumer clustering. The process involves clustering consumers using contrastive learning and k-means, training local models, and aggregating them at the server. Unlike traditional federated learning, two types of aggregation are performed: one for a global model and another for local models within clusters. Experimental validation on MovieLens-1M demonstrates the method's superiority with an average accuracy of 0.27, surpassing baseline methods capped at 0.25.