PFL Research Questions

[appleiphonedddd]

Hi

1.About the Pathological Heterogeneous Setting:

python generate_MNIST.py noniid - pat

I noticed that you mention a "pathological heterogeneous setting." Could you kindly explain how data is distributed across clients under this setting? How does it differ from the more commonly used Dirichlet distribution (often referred to as a "practical heterogeneous setting") that aims to mimic real-world non-IID data? Am I correct in understanding that the pathological setting is designed to stress-test generalization under more extreme heterogeneity?

2.On Choosing Baseline Methods for Personalized FL:
As I am currently developing a new personalized FL method based on a model-splitting approach, I am unsure about the most appropriate set of baselines for comparison. Should one always include standard federated methods (e.g., FedAvg, FedProx) alongside personalized ones? More importantly, would it be more meaningful to compare primarily against methods within the same methodological family (e.g., FedRoD, FedBABU, FedCP for model-splitting), especially if the problem statement specifically focuses on aspects like reducing communication or improving personalization?
In other words, should the choice of baseline methods depend primarily on the technical characteristics of the proposed approach, or more so on the shared problem being addressed?

Thanks you.