Identify which training-pipeline components cause federated–centralized performance differences

Identify whether federated optimization, client-specific preprocessing, partial client participation, aggregation, or another component of the respective end-to-end training pipelines generates the observed differences between FedNova and pooled centralized income estimation.

Background

The pooled centralized and federated models differ in more than their aggregation procedures. The centralized pipeline uses pooled preprocessing statistics, whereas the federated pipeline uses client-specific preprocessing; the procedures also differ in local updating, participation, and aggregation.

Although FedNova can outperform the centralized benchmark for some client groups, the reported comparisons do not isolate the contribution of these individual components. Determining the responsible mechanism would require controlled experiments that hold the remaining aspects of the pipelines fixed.