Leveraging Explainable AI for Reducing Queries of Performance Indicators in Open RAN

Open Radio Access Network (O-RAN) is positioned to play a pivotal role in shaping the future of telecommunications networks through open interfaces and virtualization, allowing interoperability between different vendors. As a key departure from single-operator managed RAN, a remote RAN intelligence controller (RIC) queries the gNB for the Key Performance Indicators (KPIs) that are required for making RAN control decisions, often leveraging advanced machine learning (ML) models. However, this repeated querying increases control traffic overhead on the so called E2 interface connecting the gNB to the RIC. To address this challenge, we utilize a method from Explainable Artificial Intelligence (XAI), specifically SHapley Additive exPlanations (SHAP), which quantifies the contribution of each requested KPI to a model's prediction. Furthermore, we explore two different methods of choosing the most discriminative KPIs influencing model's performance, so that a smaller subset of KPIs may be queried, thus lowering the overhead on the E2 interface. Our analysis reveals that a model trained for the task of traffic classification using as input only the fraction of the top contributing KPIs identified by SHAP reduces control traffic overhead by up to 33% with only 7% reduction in ML classification accuracy.