Scholarlink

<p>Software defect prediction (SDP) is considered a dynamic research problem and is&#xa0;beneficial during the testing stage of the software development life cycle. Several artificial intelligence-based methods were available to predict these software defects. However, the detection accuracy is still low due to imbalanced datasets, poor feature learning, and tuning of the model's parameters. This paper proposes a novel attention-included Deep Learning (DL) model for SDP with effective feature learning and dimensionality reduction mechanisms. The system mainly comprises ‘6’ phases: dataset balancing, source code parsing, word embedding, feature extraction, dimensionality reduction, and classification. First, dataset balancing&#xa0;was performed&#xa0;using the density peak based k-means clustering (DPKMC) algorithm, which prevents the model from having biased outcomes. Then, the system parses the source code into abstract syntax trees (ASTs) that capture the&#xa0;structure and relationship between different&#xa0;elements of the code&#xa0;to enable type checking and the representative nodes on ASTs are selected to form token vectors. Then, we use bidirectional encoder representations from transformers (BERT), which converts the token vectors into numerical vectors and extracts semantic features from the data. We then input the embedded vectors to multi-head attention incorporated bidirectional gated recurrent unit (MHBGRU) for contextual feature learning. After that, the dimensionality reduction&#xa0;is performed&#xa0;using kernel principal component analysis (KPCA), which transforms the higher dimensional data into lower dimensions and removes irrelevant features. Finally, the system used a deep, fully connected network-based SoftMax layer for defect prediction, in which the cross-entropy loss&#xa0;is utilized&#xa0;to minimize the prediction loss. The experiments on the National Aeronautics and Space Administration (NASA) and AEEEM show that the system achieves better outcomes than the existing state-of-the-art models for SDP.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel attention based deep learning model for software defect prediction with bidirectional word embedding system

  • 期刊论文

  • 发表日期：

    2025年2月1日

查看全文
  • M. Chitra Devi,
  • T. Dhiliphan Rajkumar

摘要

Software defect prediction (SDP) is considered a dynamic research problem and is beneficial during the testing stage of the software development life cycle. Several artificial intelligence-based methods were available to predict these software defects. However, the detection accuracy is still low due to imbalanced datasets, poor feature learning, and tuning of the model's parameters. This paper proposes a novel attention-included Deep Learning (DL) model for SDP with effective feature learning and dimensionality reduction mechanisms. The system mainly comprises ‘6’ phases: dataset balancing, source code parsing, word embedding, feature extraction, dimensionality reduction, and classification. First, dataset balancing was performed using the density peak based k-means clustering (DPKMC) algorithm, which prevents the model from having biased outcomes. Then, the system parses the source code into abstract syntax trees (ASTs) that capture the structure and relationship between different elements of the code to enable type checking and the representative nodes on ASTs are selected to form token vectors. Then, we use bidirectional encoder representations from transformers (BERT), which converts the token vectors into numerical vectors and extracts semantic features from the data. We then input the embedded vectors to multi-head attention incorporated bidirectional gated recurrent unit (MHBGRU) for contextual feature learning. After that, the dimensionality reduction is performed using kernel principal component analysis (KPCA), which transforms the higher dimensional data into lower dimensions and removes irrelevant features. Finally, the system used a deep, fully connected network-based SoftMax layer for defect prediction, in which the cross-entropy loss is utilized to minimize the prediction loss. The experiments on the National Aeronautics and Space Administration (NASA) and AEEEM show that the system achieves better outcomes than the existing state-of-the-art models for SDP.