Data Labeling Method in Deep Learning Model for User Clustering in the NOMA Systems

Deep Learning has emerged as a promising candidate for future wireless networks because of its strong adaptability to dynamic environments. In particular, Deep Learning algorithms are capable of learning optimal outcomes and making rapid decisions in complex scenarios. This study addresses the user clustering problem in Non-Orthogonal Multiple Access (NOMA) systems by leveraging Deep Learning techniques to maximize the system's sum rate. We initially employ a bruteforce search algorithm to explore all possible clustering strategies and identify the optimal one. This resulting dataset serves as the training input for Deep Neural Network (DNN) models. However, the generated data are often highly imbalanced, as they tend to overrepresent a limited number of high-frequency scenarios. To address this issue, the study proposed a two-step balanced sampling. First, infrequent scenarios are grouped into a single class. Second, under-sampling is applied to balance the dataset across all classes. The simulation results demonstrate that the proposed method significantly improves learning performance.