Predicting task outcome using ensemble machine learning models

Abstract

One of the most prominent problems cloud service providers face is cloud failure. The main strategy for solving this problem is fault-tolerance management, and a key method for preventing failures is failure prediction. Building a highly accurate predictive model is the primary challenge in failure prediction. In this study, we present machine learning models for predicting both multi-event (five-class) and binary task failures, using Google Cluster Traces from 2011 and 2019. We tested three feature selection methods (SelectKBest, RFE, and Random Forest importance), used PSO-based hyperparameter tuning, and applied a stacking ensemble method. We measured model performance with several metrics. Our results show that the stacking method is consistent across feature sets ( F 1 ≈ 95 % on GCT-2019), but it does not always outperform a single Decision Tree. Testing over eight years shows the model is robust to changes in workload, scheduler policies, and hardware. These results offer a starting point for predicting multi-event task termination in cloud environments.