Exploratory Data Analysis of Live 5G Radio Access Network Configuration Data Using Interpretable Machine Learning

Abstract: In the increasingly connected world of today, people are more reliant on online services than ever before. To enhance network performance, efforts needs to be made to better understand the telecommunication system. This thesis analyzes live radio access network data collected globally from multiple telecommunication operators with the objective of extracting insights and comprehend the impact of configuration parameter settings on key performance indicators. This is done using exploratory data analysis and interpretable machine learning. The work explores k-means clustering, evaluates the predictive capabilities of various machine learning models for selected key performance indicators, and uses SHAP to provide interpretability to model outputs. The findings show that k-means clustering provides insights into operators' shared parameter settings, yet it falls short of revealing the specific influential features for any of the key performance indicators. Among the machine learning models examined, tree-based approaches display the best predictive capabilities, particularly in generalizing to previously unseen operators. However, the predictability of key performance indicators varies, suggesting the influence of external factors beyond configuration parameter settings. Interpreting the machine learning model using SHAP gives explanations regarding feature impact, enabling interpretation even for individuals without expertise in machine learning. This work contributes to the research within network data analysis by reaffirming the strong performance of tree-based models shown in previous work and showcasing the potential of interpretable machine learning techniques.