A Lightweight Deep Learning Approach for Latency Prediction in 5G and Beyond

The evolution of 5 G and beyond networks is increasingly focused on enhancing Quality of Service (QoS) for end-users. Also, the advent of 5 G Multi Access Edge Computing (5 G MEC) enables enhanced optimization and service delivery leveraging the deployment of Artificial Intelligence (AI) models at the edge. Among critical QoS parameters, latency, measured through Round Trip Time (RTT), plays a crucial role, directly affecting applications such as autonomous driving and real-time services. This paper explores the adoption of a lightweight Deep Learning (DL) model for RTT prediction tailored for resourceconstrained environments at the network edge. For training and evaluation, we utilize a novel dataset collected with a smartphone-based methodology in a real-world 5G NSA scenario. Our experimental results demonstrate that a lightweight MultiLayer Perceptron (MLP) model outperforms more resourceintensive models, achieving better Mean Squared Error (MSE) compared to several baselines, making it a promising solution for edge network deployments. Eventually, we leverage the Shapley technique to enhance the interpretability of our results and the trustworthiness of the prediction, promoting the adoption of such a model in real-world scenarios.