<![CDATA[Heterogeneous Internet of Things (IoT) deployments generate highly variable network traffic, yet most existing bandwidth management approaches rely on static provisioning or device-agnostic models that fail to capture per-device traffic heterogeneity. This study proposes a hybrid machine learning framework combining K-Means clustering and Random Forest regression to enable cluster-aware, leakage-free bandwidth consumption prediction for IoT environments. Using 994,145 flow records from the UNSW HomeNet dataset, the Log1p transformation was applied to address heavy-tailed traffic distributions, and direct bandwidth-derived features were strictly excluded to prevent data leakage. Devices were segmented into three traffic profiles (Low, Medium, and High), validated by a Silhouette Score of 0.4153 and a Davies–Bouldin Index of 1.0357. Within the structural constraints of this single-dataset, controlled setting, Random Forest regression achieved R² values above 0.99 across all clusters, with stable cross-validated performance, substantially outperforming Linear Regression, which achieved R² values as low as 0.0374 in the Medium cluster. Feature importance analysis revealed that Inter-Arrival Time is the primary driver of bandwidth consumption, superseding raw payload volume, with direct implications for temporally-driven feature design. These results indicate that cluster-aware, leakage-free models offer a viable architecture for IoT bandwidth management, with cross-domain generalization as a primary direction for future work.]]>
