Contribution-Aware Personalized Federated Learning for Consumer Electronics Products Trustworthy Recommendation

Due to the development and wildly-used scenarios of consumer electronics products, it is important to build a recommendation model to select a suitable product without exposing user privacy information. Federated learning is a promising approach to providing personalized trustworthy recommendations. Its performance is inseparable from the high-quality user behavior data contributed by each participant, hence the contribution evaluation is crucial. Shapley Value is widely used in cooperative games to calculate the individual marginal contribution, but the ideal calculation method is very expensive and time-consuming in federated learning. In this paper, we propose a contribution-aware personalized federated learning named CA-PFL for consumer electronics products trustworthy recommendation. It efficiently evaluates participant contributions based on the gradient without retraining federated sub-models. Firstly, taking the Shapley value as the contribution indicator, the proposed gradient approximate estimation method reduces the original exponential time complexity to linear calculation. In addition, CA-PFL obsoletes participants with extremely low contributions and trains personalized models with different feature preferences. Comprehensive experiments are conducted on both classical machine learning datasets and consumer behavior datasets, the results demonstrate the better performance of CA-PFL from accuracy and efficiency aspects.