Hybrid feature selection module for improving performance of software vulnerability severity prediction model on textual dataset

Abstract

Software vulnerability severity prediction is a critical area in software engineering, where model performance heavily depends on the quality of the feature set used for training. Challenges such as feature redundancy, correlations, and irrelevant features can degrade model effectiveness, emphasizing the importance of Feature Selection (FS) methods to optimize performance and reduce development costs. In this study, we introduce two innovative FS modules within the homogeneous wrapper method category. The first, Parallel-Grey Wolf Optimization (P-GWO), employs a hybrid approach combining Grey Wolf Optimization (GWO) with Opposition-Based Learning (OBL). The second, Multi-Stage Grey Wolf Whale Optimization (MS-G2WO), uses GWO to find an initial optimal solution, which is further refined by the Whale Optimization Algorithm (WOA). Both modules are evaluated using Area Under Curve (AUC) values, demonstrating the significant impact of FS on model performance. Our experimental results show that P-GWO achieved superior performance with a mean AUC of 0.804, followed by MS-G2WO with a mean AUC of 0.77, establishing the effectiveness of these proposed methods in improving vulnerability severity prediction models