Development of a Federated Learning and Knowledge Graph-Based Personalized E-Commerce Recommendation System Using PTB-MFA

Abstract

E-commerce platforms require accurate and personalized recommendation systems while ensuring user privacy and scalability. Traditional centralized recommendation models face challenges related to privacy risks, data heterogeneity, and limited personalization. To address these limitations, this study proposes a Painting Training-Based Optimization with Modified Federated Averaging (PTB- MFA) framework that integrates federated learning with knowledge graph embeddings for privacy- preserving personalized recommendation. The proposed method captures complex semantic relationships among users, products, and categories while enabling adaptive client-level personalization. Experiments conducted on an e-commerce personalized recommendation dataset demonstrate that PTB-MFA outperforms existing methods, achieving improvements of 16.9% in Click-Through Rate, 19.9% in user engagement, and 6.37% in Mean Average Precision. The results confirm that the proposed framework effectively balances personalization accuracy, privacy preservation, and computational efficiency for next-generation e-commerce systems.