Enerji hasadı ve heterojen veri dağılımı koşullarında kablosuz kanallarda havadan beslemeli federe öğrenme

With the growing interest in machine learning (ML), federated learning (FL) has emerged as a prominent paradigm for collaboratively training high-quality models across decentralized edge devices, while minimizing server-side data access and preserving user privacy. In FL, multiple mobile devices (MDs) are used to collaboratively train a global model with coordination from a parameter server (PS), with devices using their local data to perform stochastic gradient descent (SGD) for the purpose of global training. While extensive research has been conducted on FL over wireless channels and its applications in practical scenarios, the need to investigate and develop solutions tailored to highly heterogeneous setups is an open research direction. This thesis aims to explore strategies and propose solutions for over-the-air (OTA) FL over wireless channels in such heterogeneous environments. In the first part of this thesis, we focus on the scheduling strategies for the OTA FL with energy harvesting MDs and highly heterogeneous data distribution. We follow a paradigm to schedule users based on their data distribution characteristics and subsequently their transmitted updates, with the primary goal of improving the learning performance and making more efficient use of the limited harvested energy. We develop two scheduling approaches depending on whether the users' data distributions are known or unknown at the server. We provide a theoretical convergence analysis of the proposed OTA FL setup, which is also used to design scheduling strategies, and validate our findings through experimental analysis, demonstrating that scheduling strategies based on user characteristics can significantly enhance global learning performance and reduce redundancy in the system. In the second part of the thesis, we study clustered federated learning (CFL) approaches in OTA FL setup, focusing on personalized global models depending on users' data characteristics for energy harvesting MDs. In this part, we assume that users are naturally partitioned into distinct clusters; for instance, each cluster may consist of users interested in different categories of sports. The primary goal is to train a dedicated model for each cluster to better capture their specific preferences, while simultaneously serving all the clusters using a single parameter server through over-the-air transmission to ensure communication efficiency. To enable the simultaneous serving of each cluster, we propose different combining methods at the server side, each designed for a specific level of available channel state information (CSI). Numerical results show that with a sufficient number of receiver antennas, it is possible to support simultaneous update transmissions from multiple clusters and achieve more specialized global models with improved performance.