Reliable mmWave Blockage Detection Under Severe Imbalance: RUS/SMOTE with Ensemble Models

Millimeter-wave (mmWave) communication systems are highly vulnerable to blockage, leading to abrupt link degradation and service interruptions. Reliable blockage detection is therefore essential for enabling proactive beam management and maintaining link robustness. However, real-world mmWave blockage datasets are inherently imbalanced, leading to biased learning and poor blockage detection performance when conventional machine learning models are applied directly. In this paper, we present a comprehensive performance evaluation of machine learning-based blockage classification across different data-balancing strategies. Decision Tree (DT) and Multi-Layer Perceptron (MLP) classifiers are evaluated using the original unbalanced dataset as well as balanced datasets generated via Random Under-Sampling (RUS) and the Synthetic Minority Oversampling Technique (SMOTE). Simulation results demonstrate that data imbalance severely degrades blockage recall, resulting in 91 missed blockage events in the unbalanced case. Applying SMOTE reduces false negatives to 69, while RUS further decreases them to 36. Moreover, the MLP model trained with SMOTE achieves the best overall performance, attaining a recall of 0.932 and an ROC-AUC of 0.958, significantly enhancing blockage sensitivity and classification robustness. These findings highlight the importance of combining data-balancing strategies with advanced learning models to enable reliable, proactive blockage detection in mmWave and future 6G wireless systems.