Clustering techniques for software product line feature identification

A Software Product Line (SPL) represents a family of products in one domain. It is often specified through a feature model representing the end-user system characteristics. Most proposed feature model construction methods examine a set of product variants code to identify common blocks of elements each of which represents a feature. In this identification process, they adopt different similarity measures and clustering techniques. Consequently, they differ in the granularity and cohesiveness of the identified features, which in turn affect the structure of the constructed feature model. Evidently, this difference also impacts the reuse of the SPL when deriving a new product. This paper compares widely adopted textual clustering techniques in terms of the recall and precision in their identification of the features. More specifically, it experimentally examines Latent Semantic Indexing (LSI), Principal Component Analysis (PCA), Support Vector Machine (SVM)and Term Frequency-Inverse Document Frequency (TF-IDF). The experimental results showed that LSI has better performance than the other textual similarity techniques with an average precision rate of 68.6