Feature Terms Prediction

In Software Product Lines (SPL), feature extraction from software requirements specifications has been subject to intense research in order to assist domain analysis in a time-saving way. Although various approaches are proposed to extract features, there still exists a gap to achieve the complete view of features, that is, how to figure out the intention of a feature. Feature terms as the smallest units in a feature can be regarded as vital indicators for describing a feature. Automated feature term extraction can provide key information regarding the intention of a feature, which improves the efficiency of domain analysis. In this paper, we propose an approach to train prediction models by using machine learning techniques to identify feature terms. To this end, we extract candidate terms from requirement specifications in one domain and take six attributes of each term into account to create a labeled dataset. Subsequently, we apply seven commonly used machine algorithms to train prediction models on the labeled dataset. We then use these prediction models to predict feature terms from the requirements belonging to the other two different domains. Our results show that (1) feature terms can be predicted with high accuracy of ≈ 90% within a domain (2) prediction across domains leads to a decreased but still good accuracy (≈ 80%), and (3) machine learning algorithms perform differently.