Federated Learning Algorithm Handling Missing Attributes
Keiichiro Oishi, Yuichi Sei, Yasuyuki Tahara, Akihiko Ohsuga · 2023
Federated learning has gained considerable attention as the solution for handling distributed and privacy-sensitive data. This includes scenarios such as predicting conversion history on smartphones or utilizing electronic medical records held by hospitals within the machine learning context. Essentially, federated learning is a form of decentralized machine learning, where data is distributed across multiple clients/organizations. This approach eliminates the necessity for direct data exchange with a central server. However, prevailing research assumes uniformity in data attributes held by each organization, disregarding the missingness attribute. As a result, this method falls short when it comes to effectively learning from data that exhibit missingness. To address this limitation, our study introduces a federated learning algorithm designed to achieve high accuracy even when dealing with data that demonstrate missingness. This algorithm incorporates a machine learning model tailored for attribute imputation within a federated learning framework. Through a series of validation experiments, we showcase the algorithm's effectiveness.