Study on the key performance of 5G downlink air interface based on an explainable machine learning model

The cell downlink effective throughput and user plane latency are important performance indicators of the 5G system’s air interface. To study the impact factors behind them, the integration of Shapley value with local linear forest model was used to make predictions with explanations. The experiment results show that compared to traditional method, this method achieves a better predict precision. Based on this, by calculating each factor’s contribution towards the results and analyzing the corresponding trends, a problematic factor for each problematic cell was found. Finally, by sorting the importance of each factor’s contribution, changes in results were inferred by changes in important factors, thus more comprehensive understanding and analysis of the performance indicators were achieved.