Multi-Class Failure Prediction for Distributed Systems Based on KPI Data and Feature Selection

In the field of distributed system failure prediction, most existing research only focuses on specific categories of failure prediction and cannot predict multiple types of failures simultaneously. In this article, we propose a novel multi-class failure prediction method KFPCIFS. KFPCIFS is built based on historical KPI metrics data. Due to not all KPI metrics being necessarily relevant to a specific type of failure, we use feature selection methods to improve the multi-class OvR strategy in KFPCIFS. Because of the limited number of failure cases, the dataset is often imbalanced. In KFPCIFS, we use the SMOTE method to address the issue of class imbalance in the dataset. Empirical studies on the dataset show the effective of KFPCIFS after comparing with eleven baselines.