Scholarlink

<p>Software defect prediction plays a crucial role as a quality assurance technology in software development. The software metrics are associated with the software quality and are vital for prediction models. Most existing defect prediction methods build the prediction model ignoring the complementary information between these two kinds of metrics. In this work, we intend to jointly leverage these two kinds of metrics. For a software instance, we regard the product metrics and the process metrics as its two views. We model the problem of discriminative feature learning from these two kinds of metrics as the problem of multi-view learning. However, it is a challenging task to construct an effective prediction model based on both product and process metrics due to the heterogeneity in data of product and process metrics, and the defect data often has class imbalance characteristic. How to explore the discriminant both inter-view and intra-view effectively has not been well studied. These characteristics make it challenging to construct an effective prediction model. In this paper, we propose a Deep Multi-view Defect Prediction (DMDP) approach, which can predict software defect based on both product and process metrics. We design a neural network with two sub-network branches, which are enforced to share the weights in the last output layer, to map the data from different views to a common space. To guide the training of networks, we design the loss function including the discrepancy loss, discrimination loss and classification loss, which further promotes the distribution consistency across views, makes full use of label information to obtain the discriminative representations, and utilizes the complementarity information for prediction. To alleviate the class imbalance problem, we design a dynamic sampling strategy for dealing with class-imbalanced data. Comprehensive experiments are conducted on 15 projects from three widely used defect datasets. The experimental results demonstrate that multi-view learning based on product and process metrics is helpful for software defect prediction and DMDP outperforms the state-of-the-art baselines.</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

Multi-view learning based on product and process metrics for software defect prediction

  • 期刊论文

  • 发表日期：

    2025年2月4日

查看全文
  • Ying Sun,
  • Fei Wu,
  • Di Wu,
  • Xiao-Yuan Jing,
  • Yanfei Sun

摘要

Software defect prediction plays a crucial role as a quality assurance technology in software development. The software metrics are associated with the software quality and are vital for prediction models. Most existing defect prediction methods build the prediction model ignoring the complementary information between these two kinds of metrics. In this work, we intend to jointly leverage these two kinds of metrics. For a software instance, we regard the product metrics and the process metrics as its two views. We model the problem of discriminative feature learning from these two kinds of metrics as the problem of multi-view learning. However, it is a challenging task to construct an effective prediction model based on both product and process metrics due to the heterogeneity in data of product and process metrics, and the defect data often has class imbalance characteristic. How to explore the discriminant both inter-view and intra-view effectively has not been well studied. These characteristics make it challenging to construct an effective prediction model. In this paper, we propose a Deep Multi-view Defect Prediction (DMDP) approach, which can predict software defect based on both product and process metrics. We design a neural network with two sub-network branches, which are enforced to share the weights in the last output layer, to map the data from different views to a common space. To guide the training of networks, we design the loss function including the discrepancy loss, discrimination loss and classification loss, which further promotes the distribution consistency across views, makes full use of label information to obtain the discriminative representations, and utilizes the complementarity information for prediction. To alleviate the class imbalance problem, we design a dynamic sampling strategy for dealing with class-imbalanced data. Comprehensive experiments are conducted on 15 projects from three widely used defect datasets. The experimental results demonstrate that multi-view learning based on product and process metrics is helpful for software defect prediction and DMDP outperforms the state-of-the-art baselines.