Federated Learning Assisted Interactive EDA with Dual Probabilistic Models for Personalized Search

Personalized search is essentially a qualitative optimization problem since its target is to find items (as solutions) satisfied by the searcher. Interactive evolutionary computation (IEC) is powerful in solving this problem in view of optimization. The privacy protection when using other users’ information in the personalized search, however, has not been concerned when designing IECs. We here present an improved interactive estimation of distribution algorithm (IEDA) with dual probabilistic models by integrating the Federated Learning (FL) proposed for privacy protection. The Federated-SVD is first developed by embedding the singular value decomposition (SVD)-based collaborative filtering into the structure of FL for safely gaining the social preference. The decomposed user and item (solution) features by SVD are uploaded and aggregated in the central service and finally used to construct and update the probabilistic models. The superiority of the enhanced IEDA is demonstrated through ten personalized search cases on movies and TV series.