On the Difficulties of Supervised Event Prediction based on Unbalanced Real-World Data in Multi-System Monitoring

Online failure prediction of performance-critical events is an important task in fault management of software systems. In this paper, we extend our previous multi-system event prediction by analyzing its performance on unbalanced, real-world data, which represents a realistic online scenario. We train a random forest classifier with different data preprocessing configurations, including data augmentation to cope with the extreme class imbalance. The results reveal that the prediction quality of the tested multi-system model drops significantly compared to the balanced scenario. Although our supervised event prediction approach as well as different data preprocessing configurations turned out to be ineffective, we consider the insights of our work valuable for the community.