Capsule feature selector for software defect prediction

The performance of predictive models in software defect prediction is vulnerable to redundant features. Feature selection methods prove effective in reducing the spatial dimensionality of features. However, traditional approaches in feature selection encounter challenges in efficiently delivering the optimal subset of features. We introduce a capsule feature selector (CFS) for software defect prediction, drawing inspiration from capsule networks. Initially, our method outlines four essential computational processes within the capsule feature selector: capsule structure, sorting rules, search rules, and transfer rules. Subsequently, we uniformly place the binary capsule operator into the capsule basic units, employing search rules to refine the search process. Sorting rules are applied to select the optimal binary capsule operator for each individual capsule basic unit. Lastly, the optimal binary capsule operator from each capsule basic unit is transmitted to the capsule region for fitness value comparison via the transfer rules. This process facilitates the selection of the optimal subset of features. Comparative experiments conducted on 15 publicly available software defect datasets demonstrate that CFS outperforms the other eight advanced feature selection algorithms. This indicates that CFS exhibits superior performance. The experimental findings are further substantiated through nonparametric statistical analysis.