Bias-Corrected Federated Learning for Video Recommendation over Stochastic Communication Links

With the increasing demand for privacy-preserving and real-time personalized services in large-scale video platforms, designing robust federated recommendation frameworks over practical communication networks has become increasingly important. To this end, this paper proposes a bias-corrected federated learning framework tailored for video recommendation over stochastic communication links. At the local training stage, a bias-corrected mechanism is introduced to explicitly account for video duration and user activity, mitigating feature-level bias and enabling the learned representations to more accurately reflect users' intrinsic preferences. To meet the timeliness requirements of real-time federated learning, the successful upload probability of local model transmission is analytically characterized under time-varying channel conditions. Building upon this probabilistic model, a statistically corrected global aggregation strategy is designed to preserve the unbiasedness of the global update with respect to the ideal fully reliable FedAvg scheme, even when a subset of local nodes fails to upload their models within the specified delay constraint. Comprehensive experimental evaluations validate that the proposed framework significantly improves recommendation accuracy and maintains robustness against communication unreliability in practical distributed environments.