Transformer-Based User Clustering for Efficient Downlink NOMA System

Abstract

As a result of the growing requirement for intelligent and adaptive resource allocation in future wireless networks, the growing interest in next-generation (NG) wireless networks has promoted the use of sophisticated user clustering methods within non- orthogonal multiple access (NOMA) systems. This paper proposes a novel deep learning framework based on a Transformer encoder for efficient user clustering and pairing in downlink NOMA. Instead of relying on text-based tokenization, the numerical channel state information (CSI) is mapped into dense feature embeddings, which are processed through multi-head self-attention to learn fine- grained inter-user relationships. This enables the model to capture interference patterns and contextual channel dependencies that conventional clustering approaches cannot represent. Using user distance, channel gain, SINR, and power allocation, we generated a synthetic dataset that meets the requirements of 3GPP TR 38.901 for use in evaluating performance in real-world fading conditions. We compared the performance based on a Transformer encoder approach with standard clustering methods (K-means, Balanced K-means, DBSCAN). The simulation results indicate that the proposed Transformer-based user clustering framework consistently outperformed all other clustering methods with respect to the key performance indicators of bit error rate (BER), throughput, user fairness, energy efficiency, and outage probability. For each of the SNR regimes, we achieved lower BERs, greater potential rate, better fairness indices, and less outage than the other clustering approaches. These results highlight the strong potential of Transformer-based architectures as scalable and intelligent solutions for NOMA user clustering and resource optimization in emerging 6G wireless networks.