README.md

The details of experiment

We show the implementation details of all experiments in the page.

Batch size

We use the same batch size for all FL methods in each dataset.

Dataset	batch size
eICU	256
skin cancer	128
ECG	64

Hyperparameters for FL methods

We utilized the grid search map used in the existing papers(FedProx, FedOpt, FedDyn).

• learning rate ($\eta$)
  • := [0.1, 0.01, 0.001, 0.0001] (FedAvg, FedProx, FedBn, FedPxN, FedDyn)
  • := [0.1, 0.03, 0.01, 0.003, 0.001, 0.0001] (FedAdam, FedAdagrad, FedYogi)
• mu ($\mu$) := [1.0, 0.1, 0.01, 0.001, 0.0001]
• feddyn alpha ($\alpha$) := [0.0001, 0.001, 0.01, 0.1]
• server learning rate ($\eta_{g}$) := [0.1, 0.03, 0.01, 0.003, 0.001, 0.0001]
• tau ($\gamma$) := [0.0001, 0.001, 0.01, 0.1]

For the grid search, we used 10% of total data in skin cancer & ECG and the data of 5 largest clients in eICU. You can check the used hyperparameters in the links below.

• Mortality prediction (mort_24h, mort_48h)
• Length of Stay prediction (LOS)
• Imminent discharge prediction (disch_24h, disch_48h)
• Final Acuity prediction (Acuity)
• Skin cancer classification (skin_cancer_image)
• Cardiac arrhythmia classification (ECG)