A Comparison of Client Weighting Schemes in Federated Learning

Data is the new oil of the digital economy. Many business organizations are gathering and using their data to optimize their business performance. However, in some cases, an individual organization may not have a sufficient amount of data or data quality to build a well-performing model, especially in a dynamic environment. In some cases, the companies, which collect similar data, may be willing to exchange their knowledge, yet without sharing their data, e.g. due to privacy or legal issues. In such a case, federated learning is a solution. In horizontal federated learning, each client (organization) iteratively improves its model, so that it can be regularly aggregated and shared with all clients participating in the federation for further improvements. In the aggregation mechanism based on the weighted average, weights depend on the model’s quality for each client. In this paper, we extend our previous approach to federated learning with missing information, and we investigate different weighting schemes, that depend on the effectiveness of the local models. The initial results are promising.