How does distributed data skew correction affect bias and variance in federated learning systems?

Data skew correction reduces bias from non-IID data but may increase variance due to synthetic or reweighted updates.

In federated learning systems, each client has its own local data distribution, often non-IID. This introduces bias in global model aggregation because some distributions dominate updates.

To correct skew, techniques like reweighting, normalization of gradients, and synthetic data augmentation are used. While these reduce bias by balancing contributions, they can introduce variance because synthetic or adjusted gradients may not reflect true local distributions.…