Enhancing Collaborative Recommendations with Federated Learning in E-Commerce

Objectives: This study explores the enhancement of collaborative recommendation systems in e-commerce platforms using federated learning (FL) integrated with glocalization strategies, focusing on data privacy preservation. The aim is to leverage natural language processing (NLP)-based models to capture real-world product relationships without sharing raw data across different platforms. Methods: An NLP-based recommendation system employing a transformer-based architecture forms the core of this study. The system processes tokenized product names to determine item relationships, encapsulating the complexities of consumer behavior patterns. The glocal federated learning (glocal FL) approach is implemented by selectively sharing encoder and decoder parameters, thus allowing certain layers to remain private. The study is conducted across distinct datasets, specifically those from retailer and grocery sectors, evaluated under both Homogeneous and Heterogeneous settings. This methodological approach facilitates a thorough investigation into the system’s adaptability to varying data environments. Results: The findings indicate that glocal FL significantly improves HR@20 scores in Homogeneous environments, surpassing local models by up to 10%. However, performance declines in Heterogeneous environments due to data variability, highlighting the need for refined strategies to manage cross-domain complexities. Conclusions: FL with glocalization offers a promising approach to enhance e-commerce recommendations by balancing global learning with local customization. This strategy effectively addresses data heterogeneity and supports privacy-preserving business communication. While proving effective in homogeneous settings, further adaptation is needed to optimize performance in diverse environments. This study underscores the potential of glocal FL to improve recommendation accuracy and maintain consumer data privacy in e-commerce.