Persistence of geometric-loss benefits under DBI optimization

Determine whether the positive effect of including the geometric regularization loss persists when FedDCN is optimized for the Davies–Bouldin index rather than clustering accuracy.

Background

FedDCN includes a geometric regularization term intended to align local feature spaces and prevent latent-space distortions. The paper reports that, when the model is evaluated with hyperparameters optimized for accuracy, including this term improves all reported metrics except the Davies–Bouldin index in the USPS IID and Fashion-MNIST IID scenarios.

Because the Davies–Bouldin index measures internal cluster quality and is not the objective used for hyperparameter optimization in the reported experiments, it remains unresolved whether the observed benefits of geometric regularization would persist if FedDCN were instead optimized directly for the Davies–Bouldin index. The authors identify this as a follow-up research question.