Resource Management in LoRaWAN With Interpretability

Long Range Wide Area Network (LoRaWAN) has become a practical communication layer for large-scale Internet-of-Things and industrial monitoring systems where battery-powered devices must transmit data reliably over long ranges. Selecting the right spreading factor (SF) remains critical for maintaining link quality and power efficiency, yet existing adaptive data-rate methods often fail under dynamic interference and mobility. This paper presents XSF, an interpretable machine-learning framework that predicts the optimal SF for each device using lightweight multilayer perceptrons (MLP) and majority voting. The model combines physical-layer indicators and device location to adapt transmission settings without changing the LoRaWAN protocol. In this paper, first, a one-time dataset has been generated using ns-3, and five MLP were trained on the dataset. Then a majority voting was employed to choose the best SF. Finally, we utilized SHapley Additive exPlanations (SHAP) analysis for interpretability for SF classification. Furthermore, the pre-trained model was then utilized in ns-3 on the end devices for efficient SF allocation based on newly generated data during simulation. The proposed XSF achieved an average classification accuracy of 85%, with up to 10.7% and 6.0% higher packet success ratio in mobility and static scenarios, respectively, while reducing energy consumption by 33.6% and 20.6% compared with the state-of-the-art baseline AI method.