Non-orthogonal multiple access (NOMA) is an important technology in addressing next generation wireless network demand, but its traditional Successive Interference Cancellation (SIC) detector deteriorates under sparsely scattered channels and with error propagation effects. Hence, the proposed work focusses an Attention-based Recurrent Neural Network (Att-RNN) applied to joint multi-user uplink Channel Estimation (CE) and Signal Detection (SD) to avoid such limitations. The proposed model learns long-range dependencies that are lost in regular RNNs by choosing to emphasize the most relevant temporal factors over stacked streams of symbols with the embedding of a self-attention layer inside a gated recurrent structure. The Att-RNN is tested in simulation experiments and the results indicate that the attention-based mechanism generates lower Symbol-Error Rates (SER) and greater Signal-to-Noise Ratio (SNR) robustness across various power-allocation and channel-mobility conditions. These outcomes validating the approach as a dependable, low-latency alternative to NOMA systems for 5G and beyond.