MJSTEM

Abstract

Abstract- Software Product Line Engineering (SPLE) is a paradigm for developing a family of software systems as against developing a single software system. Feature Model (FM) has been a defacto tool for capturing and presenting the variables and common characteristics of the entire product line, in the form of discriminating features. A Feature Model can become too complex, which makes it challenging to maintain. Consequently, this study follows up the application machine learning model in predicting the maintainability of a feature model. We evaluate the impact of feature selection, ensemble learning, and label noise correction on predictive power of machine learning models. The results show that feature selection techniques can significantly improve model efficiency. Ensemble methods yield mixed results, while label noise correction substantially enhances model performance. We identify the Number of Features (NF), Number of Leaf Features (NLeaf), and Maximum Tree Depth (DTMax) as the most influential attributes in determining feature model maintainability. In particular, a simple Decision Tree model trained on a dataset free of label-noise and with only the three most influential features mentioned above, achieves the highest accuracy (92.45%) and F1 score (93%).

Keywords