Few Features are Enough: Communication-Efficient AI-RAN

The deployment of Artificial Intelligence and Machine Learning (AI/ML) in AI-Radio Access Networks (AI-RAN) should balance the performance with communication load over the interface and model complexity. We propose a communication-efficient feature selection framework fully compliant with the O-RAN architecture. Key Performance Indicator (KPI) traffic in the non-RT RIC is analyzed to quantify the statistical relevance of each KPI to the task. A compact feature mask is then deployed to the Distributed Unit, enabling real-time filtering so that only the most informative KPIs are transmitted to the near-RT RIC for online inference. Using realistic O-RAN compliant datasets, we evaluate the proposed framework with classifiers for the slice classification use case. Results show that our framework sustains high performance while significantly reducing the number of KPIs, model parameters, and communication overhead, thereby demonstrating its suitability for scalable, low-latency AI-RAN deployments.