XAI-Driven Green Cloud Intelligence for Transparent and Optimized Resource Allocation in Cloud Computing

Abstract

The energy footprint of hyperscale cloud infrastructure has become a first-order engineering and environmental concern, yet the machine-learning controllers that increasingly govern resource allocation remain opaque, which discourages their adoption by operators who must justify every consolidation or migration decision. This paper introduces Green Cloud Intelligence (GCI), an explainable framework that couples an optimised gradientboosted decision engine with a Shapley-value transparency layer and an energy-aware allocation policy, so that each placement decision is both accurate and accountable. The framework ingests multivariate host telemetry, applies a hybrid feature-selection stage that combines mutual information with SHAP-gain ranking, and trains a classbalanced, hyperparameter-tuned XGBoost classifier that maps each host-hour to one of three operational decisions— consolidate, maintain, or relieve. A physically grounded power and demand model is used to generate a reproducible twelve-feature workload corpus of 12,000 host-hour records against which the framework and seven representative baselines are evaluated. The proposed model attains 95.06% accuracy, a macro-F1 of 0.9387, and a macro one-vsrest ROC-AUC of 0.9947, outperforming logistic regression, decision tree, k-NN, SVM, random forest, and a multilayer perceptron while retaining sub-millisecond inference. When its decisions are translated into a consolidation policy, GCI realises a net data-centre energy reduction of 19.45%, capturing roughly 92% of an oracle upper bound. SHAP attribution shows the decisions are driven by physically meaningful signals—processor utilisation, virtual-machine density, and inlet temperature—rather than spurious correlations, and an ablation study confirms that a compact eight-feature model preserves accuracy at reduced cost. The novelty lies in unifying predictive accuracy, native transparency, and quantified energy accountability within a single deployable pipeline. Future work will validate the framework on public production traces and integrate carbon-intensity forecasting for temporally shifted scheduling