Modeling Relevance Relations Using Machine Learning Techniques

Abstract

This chapter presents the notion of relevance relations, an abstraction to represent relationships between software entities. Relevance relations map tuples of software entities to values that reflect how related they are to each other. Although there are no clear definitions for these relationships, software engineers can typically identify instances of these complex relationships. We show how a classifier can model a relevance relation. We also present the process of creating such models by using data mining and machine learning techniques. In a case study, we applied this process to a large legacy system; our system learned models of a relevance relation that predict whether a change in one file may require a change in another file. Our empirical evaluation shows that the predictive quality of such models makes them a viable choice for field deployment. We also show how by assigning different misclassification costs such models can be tuned to meet the needs of the user in terms of their precision and recall.