Software defect prediction aims to identify the potential defects of new software modules in advance by constructing an effective prediction model. However, the model performance is susceptible to irrelevant and redundant features. In addition, previous studies mainly use traditional data mining or machine learning techniques for defect prediction, the prediction performance is not superior enough. For the first issue, motivated by the idea of search based software engineering, we leverage the recently proposed whale optimization algorithm (WOA) and another complementary simulated annealing (SA) to construct an enhanced metaheuristic search based feature selection algorithm named EMWS, which can effectively select fewer but closely related representative features. For the second issue, we employ a hybrid deep neural network — convolutional neural network (CNN) and kernel extreme learning machine (KELM) to construct a unified defect prediction predictor called WSHCKE, which can further integrate the selected features into the abstract deep semantic features by CNN and boost the prediction performance by taking full advantage of the strong classification capacity of KELM. We conduct extensive experiments for feature selection or extraction and defect prediction across 20 widely-studied software projects on four evaluation indicators. Experimental results demonstrate the superiority of EMWS and WSHCKE.

软件缺陷预测旨在通过构建有效的预测模型，提前识别新软件模块的潜在缺陷。然而，模型性能容易受到不相关和冗余特征的影响。此外，以往的研究主要采用传统的数据挖掘或机器学习技术进行缺陷预测，预测性能不够优越。对于第一个问题，受基于搜索的软件工程思想的启发，我们利用最近提出的鲸鱼优化算法 (WOA) 和另一种互补模拟退火 (SA) 构建了一种名为 EMWS 的基于增强元启发式搜索的特征选择算法，该算法可以有效地选择较少但密切相关的代表性特征。对于第二个问题，我们采用混合深度神经网络——卷积神经网络 (CNN) 和内核极限学习机 (KELM) 构建了一个统一的缺陷预测预测器，称为 WSHCKE，它可以进一步将所选特征集成到 CNN 抽象的深度语义特征中，并提升通过充分利用 KELM 强大的分类能力来提高预测性能。我们对 20 个广泛研究的软件项目的四个评估指标进行了广泛的特征选择或提取和缺陷预测实验。实验结果证明了 EMWS 和 WSHCKE 的优越性。可以进一步将选择的特征融合到 CNN 抽象的深层语义特征中，充分利用 KELM 强大的分类能力，提升预测性能。我们对 20 个广泛研究的软件项目的四个评估指标进行了广泛的特征选择或提取和缺陷预测实验。实验结果证明了 EMWS 和 WSHCKE 的优越性。可以进一步将选择的特征融合到 CNN 抽象的深层语义特征中，充分利用 KELM 强大的分类能力，提升预测性能。我们对 20 个广泛研究的软件项目的四个评估指标进行了广泛的特征选择或提取和缺陷预测实验。实验结果证明了 EMWS 和 WSHCKE 的优越性。