摘要
软件缺陷预测是提升软件质量与降低维护成本的重要手段，随着人工智能技术的快速发展，其在软件工程领域的应用日益广泛。本研究旨在构建一种基于人工智能的软件缺陷预测模型，以提高预测精度和实用性。为此，我们提出了一种融合深度学习与特征选择算法的新型框架，该框架首先通过变分自编码器对高维数据进行降维处理，提取关键特征，随后利用长短期记忆网络（LSTM）捕捉代码变更的时间序列特性，从而实现对软件模块缺陷的精准预测。实验采用来自开源项目的多组大规模数据集，并与传统机器学习方法及现有深度学习模型进行对比分析。结果表明，所提出的模型在F1分数、AUC值等指标上显著优于其他方法，特别是在小样本场景下仍能保持较高的预测性能。此外，本研究还探讨了不同特征组合对模型效果的影响，验证了特征选择策略的有效性。总体而言，本研究不仅为软件缺陷预测提供了新的思路，还为人工智能技术在软件工程中的深入应用奠定了基础，其创新点在于结合了无监督特征提取与时间依赖建模，为实际开发过程中的缺陷管理提供了有力支持。
关键词：软件缺陷预测；深度学习；特征选择；变分自编码器；长短期记忆网络
Abstract
Software defect prediction is a critical approach for enhancing software quality and reducing maintenance costs, and with the rapid advancement of artificial intelligence technologies, its applications in software engineering have become increasingly extensive. This study aims to construct an artificial intelligence-based software defect prediction model to improve prediction accuracy and practicality. To achieve this, we propose a novel fr amework that integrates deep learning with feature selection algorithms. This fr amework first applies variational autoencoders to reduce the dimensionality of high-dimensional data and extract key features, followed by leveraging long short-term memory networks (LSTM) to capture the time-series characteristics of code changes, thereby enabling precise prediction of defects in software modules. Experiments were conducted using multiple large-scale datasets sourced from open-source projects and involved comparative analyses with traditional machine learning methods and existing deep learning models. The results demonstrate that the proposed model significantly outperforms other methods in terms of F1 score, AUC value, and other metrics, particularly maintaining high prediction performance even in small-sample scenarios. Additionally, this study investigates the impact of different feature combinations on model effectiveness, validating the efficacy of the feature selection strategy. Overall, this research not only provides new insights into software defect prediction but also lays a foundation for the deeper application of artificial intelligence technologies in software engineering. Its innovation lies in the combination of unsupervised feature extraction and time-dependency modeling, offering strong support for defect management in actual development processes.
Keywords：Software Defect Prediction; Deep Learning; Feature Selection; Variational Autoencoder; Long Short-Term Memory Network
目 录
摘要 I
Abstract II
一、绪论 1
(一) 软件缺陷预测的研究背景与意义 1
(二) 人工智能在软件缺陷预测中的研究现状 1
(三) 本文研究方法与技术路线 2
二、数据处理与特征选择 2
(一) 软件缺陷数据的获取与预处理 2
(二) 特征工程在缺陷预测中的应用 3
(三) 关键特征的选择与优化方法 3
三、模型构建与算法分析 4
(一) 基于人工智能的模型设计原则 4
(二) 常用机器学习算法的对比分析 5
(三) 深度学习在缺陷预测中的优势与挑战 5
四、实验验证与结果分析 6
(一) 实验环境与数据集介绍 6
(二) 模型性能评估指标体系 6
(三) 实验结果分析与改进策略 7
结 论 8
参考文献 9
