Federated Similarity-Based Learning with Incomplete Data
Barbara Pȩkala, Jarosław Szkoła, Krzysztof Dyczkowski, Anna Wilbik · 2023
In the analysis of social, medical, and business issues, the problem of incomplete data often arises. In addition, in situations where privacy policy makes it difficult to share data with organizations conducting related activities, it is necessary to exchange knowledge instead of data, that is, to use federated learning. In this scenario there are several private data clients, whose models are improved through the aggregation of model components. Here, we propose a methodology for training local models to deal well with missing data, with an algorithm using similarity measures that take into account the uncertainty present in many types of data, such as medical data. Therefore, this paper describes a federated learning model capable of processing imprecise and missing data. Federated learning is a technique to overcome limitations resulting from data governance and privacy by training algorithms without exchanging the data itself. The performance of the proposed method is demonstrated using medical data on breast cancer cases. Results for different data loss scenarios and corresponding measures of classification quality are presented and discussed.