Improving distributed systems failure prediction via multi-objective feature selection and deep forest

Distributed systems failure prediction uses Key Performance Indicator (KPI) metrics to train machine learning models to identify potential system failures. Currently, this approach faces the challenge of KPI metrics redundancy. Although existing studies have used feature selection methods to address this issue, the optimal number of KPI metrics remains difficult to determine. Furthermore, existing failure prediction models have not achieved satisfactory performance. To address the aforementioned challenges, we propose a failure prediction method, MOFSDF, which leverages multi-objective optimization and the deep forest model. The method aims to achieve optimal failure prediction performance with the fewest KPI metrics. MOFSDF frames the selection of KPI metrics as a multi-objective optimization problem, where the objective is to minimize the number of KPI metrics while simultaneously maximizing the model’s performance. This optimization process is achieved through the NSGA-III algorithm. Furthermore, the integration of the deep forest model guarantees both the robustness and generalization capabilities of the failure prediction model. We use the ZTE dataset for the empirical study. The results demonstrate that MOFSDF significantly outperforms other comparative methods across multiple dimensions. Compared to nine classical models, the A c c u r a c y improvement ranges from 1.1% to 97.9%, while the M a c r o − F 1 improvement ranges from 1.1% to 179.4%. The K a p p a and H a m m i n g L o s s also show significant improvements. Moreover, the findings from the ablation studies further substantiate the rationale and effectiveness of the component configuration within MOFSDF. • A distributed system failure prediction model based on gcForest is proposed. • Feature selection is treated as a multi-objective optimization problem, and multi-objective optimization is performed based on NSGA-III. • The SMOTE method is used to address the class imbalance issue in the KPI metrics dataset. • Comparative and ablation experiments were conducted based on the real-world ZTE dataset, and the experimental results demonstrate the effectiveness of our method.