SLA-AWARE PREDICTIVE RESOURCE ALLOCATION IN 6G RADIO ACCESS NETWORKS USING CLIENT-TEMPORAL TELEMETRY LEARNING
Keywords:
6G networks, resource allocation, SLA prediction, radio access network telemetry, machine learning, network slicing, quality of service, predictive resource management, calibration, risk-aware decisioning.Abstract
Sixth-generation radio access networks are expected to support heterogeneous services whose performance requirements vary across throughput, latency, reliability, mobility, and slice-level demand. Conventional resource allocation mechanisms are often reactive because network resources are adjusted after congestion or service-level agreement (SLA) degradation has already appeared. This paper proposes an SLA-aware predictive resource allocation framework for 6G radio access networks using client-temporal telemetry learning. The proposed framework predicts whether the next 90-second client-window will satisfy the SLA and converts the predicted probability into a risk score for proactive resource prioritization. The study uses the public Wireless 6G Network Dataset of Resource Allocation, containing 30,000 telemetry observations from 200 clients. Each observation includes radio-quality, physical-resource-block usage, retransmission, mobility, traffic-slice, throughput, latency, reward, and current-SLA indicators. A client-aware chronological split is used to preserve deployment realism, producing 21,000 training, 4,400 validation, and 4,600 test observations. Temporal lag, delta, and rolling-window features are generated, resulting in 265 predictive features. Seven machine learning models are evaluated: Logistic Regression, Random Forest, Gradient Boosting, HistGradientBoosting, multilayer perceptron, XGBoost, and LightGBM. Results show that Gradient Boosting achieves the best overall predictive performance with 0.8585 accuracy, 0.8566 macro-F1, 0.9403 ROC-AUC, and 0.9573 PR-AUC. Random Forest achieves the strongest violation recall of 0.9140, reducing missed SLA violations to 167 out of 1,941 violation cases. HistGradientBoosting provides the best probability calibration with an expected calibration error of 0.0176. These findings demonstrate that calibrated telemetry learning can support predictive SLA assurance and risk-aware resource prioritization in 6G networks.
