KNAS-TP: Domain Knowledge-driven Automated Construction of Uplink and Downlink Throughput Predictor in 5G

The advent of 5G networks marks a revolutionary shift in wireless communications, offering unprecedented improvements in speed, latency, and connectivity. Accurate uplink and downlink throughput prediction is crucial for 5G network Operations and Maintenance (O&M) tasks, yet existing methods constructing intuitive predictors without structural optimization often fail to adapt to dynamic 5 G environments due to the lack of effective utilization of domain knowledge. To address this, we present KNAS-TP, the first domain knowledge-driven automated construction method for 5G throughput prediction based on Neural Architecture Search (NAS) and the Attention mechanism. KNAS-TP incorporates domain knowledge in NAS search space establishment and loss function definitions, which includes the match degree evaluation of the complexity of the searched neural network structure and corresponding knowledge. Additionally, KNAS-TP enhances model using an Attention mechanism to captures the uneven impacts of deep factors, thereby improving final model training and prediction accuracy. Extensive experimental results demonstrate that our model outperforms all comparative methods, achieving up to $8.5 \times$ lower throughput prediction MSE. The integration of domain knowledge is essential for robust and accurate throughput predictions, validating the effectiveness of our methodology in complex 5G scenarios.