A form of data heterogeneity where clients have non-uniform distributions of class labels, simulated using Dirichlet distributions.