Proactive Cloud Systems Reliability: A Multi-Model Approach to Node Failure Prediction

Cloud services call for high reliability, motivating strategies to predict server node disruption before it arises. This paper proposes a multi-model failure prediction framework that combines deep learning for temporal analysis with ensemble learning for spatial analysis of machine states. Specifically, a Transformer-based sequence model captures time-series patterns in resource usage, replacing a conventional LSTM method, and an XGBoost classifier evaluates instantaneous device capabilities. The outputs are fused and passed into a second XGBoost model that ranks nodes by failure risk. To handle extremely uncommon failure events, the proposed model applies class-balancing techniques, including focal loss during training and SMOTE oversampling, and a cost-sensitive thresholding mechanism. The cost-sensitive approach allows tuning the false negative vs. false positive trade-off via weighting their "costs" to align with operational priorities for example, heavily penalizing missed failures on critical nodes. We evaluated our method on actual cloud trace data. Results show that it can achieve 80% recall, catching most imminent failures with about 100% precision at the chosen operating point, effectively providing early warning for most failures with zero false alerts. We also compared our method to a reference solution by Microsoft Research, which uses an LSTM and a Random Forest in a similar ensemble way. Moreover, explainable AI techniques, SHAP, is employed to interpret model outputs, identifying key signals that drive the predictions. These insights enhance the transparency and trustworthiness of the model.