Modular Federated Cross-Domain Recommendation (MFCDR) with a Projected Attention Network

In today’s digital age, data is considered a new currency. It drives many aspects related to research, business strategies, and decisions across various industries, providing recommendations. As data becomes more valuable, the privacy of user data becomes crucial. This article introduces an innovative, novel privacy-preserving federated learning approach, with the advancement of attention networks for cross-domain recommendation. A decentralized approach to federated learning has replaced traditional machine learning to enhance user privacy and data security. The research employs a projected attention network (PRADO) within a real federated environment to improve local device training. The proposed framework is examined through a use case of book recommendation based on emotions extracted from social media reviews, utilizing GoEmotions and customized book datasets. The implementation leverages advanced methodologies, including the PRADO architecture and buffered FedAvg, which outperform the base models in emotion prediction across various metrics, in particular, precision, the F1 score (11% better than an average of the base models), and balanced recall, thereby validating a robust approach. Integrating the emotional context into recommendations (with a precision of 0.96) addresses complex user preferences. The research demonstrates that the real implementation of a federated learning environment is modular, scalable, efficient, and preserves privacy.