An extended study on applicability and performance of homogeneous cross-project defect prediction approaches under homogeneous cross-company effort estimation situation
An extended study on applicability and performance of homogeneous cross-project defect prediction approaches under homogeneous cross-company effort estimation situation
复制标题
DOI：
10.1007/s10664-021-10103-4
复制
发表时间：
2022-01
影响因子：
4.1
通讯作者：
S. Amasaki;Hirohisa Aman;Tomoyuki Yokogawa
中科院分区：
文献类型：
--
作者：
S. Amasaki;Hirohisa Aman;Tomoyuki Yokogawa
软件工作量估算 (SEE) 模型的研究已有数十年。对于面向数据的模型来说，严重但典型的情况之一是训练模型的数据集的可用性。跨公司软件工作量估算 (CCSEE) 是一个利用组织外部项目数据的研究主题。软件缺陷预测研究中也发现了同样的问题，并研究了跨项目缺陷预测（CPDP）。 CPDP 和 CCSEE 有着相同的动机并针对同一问题开发了方法。出现了 CPDP 方法是否适用于 CCSEE 的问题。本研究通过一项调查和一项侧重于同质方法的实证研究来探讨这个问题。我们首先检查了 CPDP 框架中提供的 24 种同类 CPDP 方法实现，发现超过一半的方法是适用的。接下来，我们利用过去两项研究的结果来检查适用的 CPDP 方法是否涵盖了过去 CPDP 研究中采取的策略。然后进行实证实验来评估适用的 CPDP 方法的估计性能。 CPDP 方法配置了一些机器学习技术（如果可用），并提供了 10 个 CCSEE 数据集配置用于评估。该结果还与过去两项研究的结果进行了比较，以发现 CPDP 和 CCSEE 之间的共性和差异。研究问题的答案表明我们的 CPDP 选择涵盖了 CPDP 方法所采取的策略。实证结果支持简单合并，无需实例加权，无需特征选择，无需数据转换，也无需简单投票。根据 CPDP 和 CCSEE 研究，这并不一定令人惊讶，但尚未在 CCSEE 情况下用 CPDP 方法进行探索。实际意义是：结合跨公司数据来进行工作量估算。
Software effort estimation (SEE) models have been studied for decades. One of serious but typical situations for data-oriented models is the availability of datasets for training models. Cross-company software effort estimation (CCSEE) is a research topic to utilize project data outside an organization. The same problem was identified in software defect prediction research, and Cross-project defect prediction (CPDP) has been studied. CPDP and CCSEE shared the motivation and developed approaches for the same problem. A question arisen that CPDP approaches were applicable to CCSEE. This study explored this question with a survey and an empirical study focusing on homogeneous approaches. We first examined 24 homogeneous CPDP approach implementations provided in a CPDP framework and found more than half of the approaches were applicable. Next, we used the results of two past studies to check whether the applicable CPDP approaches covered strategies taken in past CPDP studies. The empirical experiment was then conducted to evaluate the estimation performance of the applicable CPDP approaches. CPDP approaches were configured with some machine learning techniques if available, and ten CCSEE dataset configurations were supplied for evaluation. The result was also compared with the results of those two past studies to find the commonalities and the differences between CPDP and CCSEE. The answers to the research questions revealed the our CPDP selection covered the strategies that CPDP approaches had taken. The empirical results supported the simple merge with no instance weighting, no feature selection, no data transformation, nor the simple voting. It was not necessarily surprising according CPDP and CCSEE studies, but had not been explored with CPDP approaches under CCSEE situation. A practical implication is: Combine cross-company data to make effort estimates.