Untitled Item

The Product Line Architecture (PLA) is one of the most important artifacts of a Software Product Line (SPL). PLA design can be formulated as an interactive optimization problem with many conflicting factors. Incorporating Decision Makers' (DM) preferences during the search process may help the algorithms to find more adequate solutions for their profiles. Interactive approaches allow the DM to evaluate solutions, guiding the optimization according to their preferences. However, this brings up human fatigue problems caused by the excessive amount of interactions and solutions to evaluate. A common strategy to prevent this problem is limiting the number of interactions and solutions evaluated by the DM. Machine Learning (ML) models were also used to learn how to evaluate solutions according to the DM profile and replace them after some interactions. Feature selection performs an essential task as non-relevant and/or redundant features used to train the ML model can reduce the accuracy and comprehensibility of the hypotheses induced by ML algorithms. This work aims to select features of a ML model used to prevent human fatigue in an interactive search-based PLA design approach. We applied four selectors and through results we were able to reduce 30% of features, obtaining an accuracy of 99%.产品线架构（Product Line Architecture, PLA）是软件产品线（Software Product Line, SPL）中最为核心的软件制品之一。PLA的设计可被建模为一类存在诸多冲突因素的交互式优化问题。在搜索过程中融入决策者（Decision Makers, DM）的偏好，可帮助算法为决策者的需求画像寻得更适配的解决方案。交互式方法允许决策者对候选解决方案进行评估，并依据其偏好引导优化进程。但此类方法也会因需评估的交互次数与候选方案数量过多，引发决策者的认知疲劳问题。规避该问题的常用策略，是限制决策者需评估的交互次数与候选方案数量。机器学习（Machine Learning, ML）模型还可用于学习依据决策者的需求画像评估候选方案的方法，并在若干次交互后替代人工评估环节。特征选择（Feature selection）承担着关键作用：用于训练ML模型的无关或冗余特征，会降低机器学习算法所生成假设的准确率与可理解性。本研究旨在为一款应用于基于交互式搜索的PLA设计场景、以规避决策者认知疲劳的机器学习模型开展特征选择工作。本研究应用了四种特征选择器，实验结果表明可将特征维度缩减30%，同时保持99%的准确率。

搜集汇总

背景与挑战

背景概述

该数据集聚焦于软件产品线架构的交互式优化，通过机器学习模型减少人类疲劳，并实现了30%的特征削减和99%的准确率。数据集包含多个相关文件，适用于数据工程和科学领域的研究。