Statistical AI/ML Model Monitoring for 5G/6G: Interference Prediction Case Study

AI/ML models are anticipated to play a pivotal role in enhancing the performance of 5G/6G networks. However the inherently dynamic nature of wireless environments makes ML models susceptible to data drifts which can cause significant performance degradation. This paper evaluates statistical ML model monitoring for the interference prediction use case under data drift scenarios where a UE moves randomly between eight different environments. First, we demonstrate that a mismatch between the interference distribution in the inference environment and that used in training the ML model can significantly degrade the ML model performance which necessitates monitoring the ML models. To detect data drifts, we propose three different statistical metrics that capture statistical distances between the training data distribution and inference distribution observed in the deployment environment. We study the effect of the number of samples for the inference data distribution on the performance of data drift detection. We adapt to data drifts by switching between best-matched pre-trained ML models. Our numerical simulations demonstrate that ML model selection based on statistical monitoring enhances the interference prediction accuracy compared to different baselines.