Data Augmentation for Improving Convergence Speed in Federated Sequential Recommendation System

A federated sequential recommendation system enables personalized temporal recommendations while safeguarding user privacy. However, the statistical heterogeneity of independent user records often necessitates extensive communication to achieve high-performing models. To address this challenge, prior research in other domains has employed data augmentation techniques to mitigate heterogeneity by generating synthetic datasets. Despite their potential, data augmentations have not been systematically explored in the context of federated recommendation systems. We aim to systematically evaluate six data augmentation methods and their effectiveness in mitigating statistical heterogeneity for efficient federated sequential recommendation. Our findings indicate that augmentation techniques introducing variation in sequence lengths can enhance convergence speed and improve the generalizability of federated models, while reducing communication overhead. To our knowledge, this is the first study to systematically evaluate data augmentation in federated recommendation systems.