Interpretable KPI Analytics for Resource-Efficient AI-RAN Intelligence

Artificial Intelligence for Radio Access Networks (AI-RAN) has emerged as a critical enabler for intelligent control and optimization in next-generation wireless systems, and its realization is closely tied to the Open RAN (O-RAN) architecture. O-RAN provides the standardized, disaggregated framework within which AI-RAN intelligence can be deployed through the RAN Intelligent Controller (RIC) and its modular rApps and xApps. Within O-RAN, the RIC integrates non-real-time (non-RT) rApps and near-real-time (near-RT) xApps that rely on continuous streams of Key Performance Indicators (KPIs) generated by distributed units (DUs). KPIs are statistical measurements that capture the state of wireless traffic and radio link conditions, such as throughput, resource block allocation, and signal quality. They are central to data-driven RAN intelligence, as they provide the basis for adaptive scheduling and resource allocation. However, transmitting the full set of high-dimensional KPIs across O1 and E2 interfaces leads to communication bottlenecks, while processing them with large-scale models increases inference latency and deployment overhead. While AI/Machine Learning (ML) models have demonstrated strong performance for slice classification and other RAN control tasks, a persistent challenge lies in ensuring that the decision process remains interpretable. Interpretability is essential in operational wireless systems, as network operators must not only trust the predictions of AI/ML models but also understand which network indicators are most influential. Without interpretability, model outputs remain black-box decisions that are difficult to validate, troubleshoot, and integrate into large-scale control loops. Thus, reducing KPI dimensionality in a principled and interpretable manner is critical for scalable AI-RAN. We address this challenge with an interpretable KPI analytics framework designed for resource-efficient AI-RAN intelligence. Offline in the non-RT RIC, KPI traffic is analyzed using statistical relevance metrics, Pearson Correlation Coefficient (PCC), to quantify each KPI’s contribution to slice classification. A compact feature mask is then derived and deployed at the DU to filter incoming KPI vectors in real time. Only the most informative KPIs are transmitted to near-RT RIC xApps for classification, reducing KPI telemetry overhead and model complexity. Importantly, this approach preserves interpretability by explicitly ranking KPIs according to their relevance, enabling operators to link each retained feature to its direct impact on slice discrimination. Experiments on two O-RAN compliant datasets, COMMAG and ColO-RAN, confirm that compact KPI sets are sufficient for effective network slice classification. Across both datasets, only 7–9 out of 21 KPIs are needed to sustain performance comparable to the full set, with communication and computational overhead reduced drastically. Even with just two selected KPIs, lightweight classifiers such as XGBoost, SVM, and CNN maintain accuracy within 5% of the baseline while reducing telemetry by over 75% and shrinking model payload size by up to 90%. Beyond performance gains, interpretability analysis reveals that the most consistently selected KPIs correspond to downlink throughput and physical resource block usage, which are directly tied to slice-specific quality-of-service guarantees. This finding confirms that the discriminative KPIs are not only compact but also stable and explainable, providing meaningful insights into how slice classification decisions align with core RAN functions.