Extracting domain knowledge is important for different purposes, including development of new systems and maintenance of existing systems in the domain. Automatically supporting this task is challenging; most existing methods assume high similarity of variants which limits reuse of the generated domain artifacts, or provide very low-level features which hinder domain structure and behavior. In this paper, we propose a holistic method for extracting domain knowledge in the form of feature models that capture mandatory, optional and variant domain behaviors. Particularly, the method gets low-level implementations, applies polymorphism-inspired mechanisms and multi-criteria decision making for generating candidate domain behaviors, utilizes machine learning techniques to classify local, global and irrelevant domain behaviors, and finally analyzes dependencies and presents the outcomes in the form of feature models. The approach is evaluated on two datasets: one of open-source video games, named apo-games, following a clone-and-own scenario; and the other on variants of a monopoly game, simulating a scenario of independent development of similarly behaving components.

提取领域知识对于不同的目的很重要，包括新系统的开发和领域中现有系统的维护。自动支持这项任务具有挑战性；大多数现有方法假设变体的高度相似性，这限制了生成的域工件的重用，或者提供了阻碍域结构和行为的非常低级的特征。在本文中，我们提出了一种以特征模型的形式提取领域知识的整体方法，该模型捕获强制性、可选和可变的领域行为。特别是，该方法获得低级实现，应用多态启发机制和多标准决策来生成候选域行为，利用机器学习技术对局部、全局和不相关的域行为进行分类，最后分析依赖关系并以特征模型的形式呈现结果。该方法在两个数据集上进行评估：一个名为 apo-games 的开源视频游戏，遵循克隆和拥有的场景；另一个是垄断游戏的变体，模拟行为相似的组件的独立开发场景。