This chapter proposes a near-field beamforming design based on unsupervised learning for massive multiple-input multiple-output (MIMO) non-orthogonal multiple access (NOMA) downlink systems. In high-frequency scenarios such as millimeter-wave (mmWave), users often reside in the near-field region, where spherical wave fronts must be considered. Traditional beamforming techniques, such as zero-forcing (ZF) and minimum mean square error (MMSE), do not fully exploit the spatial characteristics of near-field channels. While recent advances have applied deep learning to beamforming, their use in near-field massive MIMO-NOMA remains limited. We introduce a convolutional neural network (CNN)-based beamforming scheme trained via unsupervised learning to maximize system sum-rate. The proposed hybrid system integrates deep learning for near-field users and successive convex approximation (SCA)-based resource allocation for far-field users. Simulation results demonstrate rapid convergence of the model, high-precision beam focusing with minimal interference leakage, and a sum-rate gain of up to 3.51 bps/Hz over ZF at an SNR of 20 dB. The proposed method ensures consistent performance improvements across the entire SNR range while maintaining minimum rate guarantees for near-field users.