Federated Learning for Rule-Based Systems: Preliminary Studies

This study proposes the novel Federated Learning for Interpretable Rule Trees (FL-IRT) method that is introduced here to address the challenges posed by non-Independent and Identically Distributed (non-IID) data distributions and model explainability for the Federated Learning (FL) framework. By exploiting Decision Tree (DT) methods at the client side for interpretability and simplicity, our method is able to accommo-date data with varying distributions across disseminated clients. Through evaluations conducted on a synthetic dataset, our approach exhibits remarkable good accuracy results and very fast convergence w.r.t. the plain Federated Averaging (FedAvg) and Federated Proximal optimization (FedProx) algorithms as well. Notably, a key advantage of our proposed algorithm, especially in IoT applications, lies in its substantially reduced computational time compared to FedAvg and FedProx, while achieving improving interpretability and explainability as well. According to our simulations, FL-IRT was 37.7% faster than FedAvg and 126% faster than FedProx. This efficiency enhancement highlights the practical viability and speed of our algorithm in the context of FL scenarios with devices with limited processing capabilities.