Abstract—Beyond 5G (B5G) networks will deploy hybrid multi-radio access technologies for improved Spectrum Efficiency (SE) and Energy Efficiency (EE). For further capacity enhancement, code and power domain multiplexing are applied on hybrid Power Domain Sparse Code Nonorthogonal Multiple Access (PD-SCMA) schemes. In order to realize optimal performance, robust Resource Allocation (RA) policies are required with the best candidate for the complex overloaded B5G systems based on Artificial Intelligence (AI) techniques. This work develops a deep neural network (DNN) aided resource allocation scheme for an uplink PD-SCMA network with Near-User (NU) and Far- User (FU) groups, multiplexed in the power and code domain. The RA problem is formulated as a non-convex joint optimization problem then decomposed into three subproblems namely Codebook Assignment (CA), User Clustering (UC) and power allocation (PA). For the three sub-problems, a three-stage generic fully connected DNN is trained to approximate the PA, CA and UC resource allocation solution by the proposed modified Primal-Dual Interior Point Method (mPD-IPM). The proposed mPD-IPM generates the near-optimal RA solutions that form the DNN input labels. The DNN-mPD-IPM not only greatly enhances the computational efficiency but also achieves improved convergence rates and guarantees both the ergodic and nonergodic sum rates of the system compared to generic algorithms. Simulation results show that the DNN aided resource allocation closely learns the system capacity and computational performance of the proposed mPD-IPM and further outperforms generic RA algorithms. Compared to cross-layer codebooks, mPD-IPM and DNN-mPD-IPM achieves approximately 19% higher capacity. The proposed DNN-mPD-IPM that learns to approximate the proposed mPD-IPM has an execution time that is approximately 70% lower than mPD-IPM.