• Explainable Artificial Intelligence
• Federated Learning
• Fuzzy Regression Trees
• Machine Learning

Federated learning (FL) is a recently emerged distributed machine learning paradigm, where more clients are allowed to collaboratively train a global Machine Learning model without sharing their private raw data, thus ensuring the preservation of privacy. Although FL has been mainly adopted in Neural Networks models, it has been recently investigated also with inherently explainable models, such as Takagi-Sugeno-Kang Fuzzy Rule-Based Systems (TSK-FRBSs) and Decision Trees.
In this thesis, we focus on a fuzzy extension of Regression Trees (RTs), namely Fuzzy Regression Trees (FRTs), which feature a high level of explainability and have proven to yield competitive performance, especially in tasks with some degree of noise and/or uncertainty. However, the classical FRT learning algorithm does not comply with the FL setting, in which raw data are spread over multiple nodes and cannot be transmitted for centralized processing. To overcome this limitation, a novel approach to build FRT in a federated fashion is proposed: the learning procedure preserves the privacy of data owners and supports polynomial approximators up to the second order in the leaf nodes.
To assess the performances of the developed approach, we carried out an extensive experimental analysis: the federated approach is compared to local learning, in which each data owner builds its own FRT based on a local dataset, and centralized learning, in which a single FRT is built after merging local datasets. We also compared it with a recently proposed version of federated TSK-FRBS on a publicly available real-world dataset for the task of Quality of Experience prediction in next-generation wireless networks.