Feature Location Using Data Mining on Existing Test-Cases

Feature location is the mapping of features onto static artifacts such as source code and configuration files. Recent effective feature location methods use dynamic-analysis and scenarios, i.e. executable use-cases that represent features in source code. Almost all of these techniques require that a distinct scenario is created for every feature, which puts a great burden on developers due to the inherent manual work that needs to be carried out. First, since scenarios are typically created and exercised manually, the act of locating features is not easily repeatable as software evolves. Second, there may be dependencies between features, which will be reflected on the collected execution traces during scenario execution. For existing feature location methods to perform well, developers typically need to mark the execution traces manually to specify which parts of it exactly represent each feature. In this work, we free developers of the manual process of creating scenarios by using association rule learning on a set of labeled test-cases, i.e. test-cases labeled with the features they exercise in the existing test-suite of the system, to locate features in source code. We also provide an evaluation of our method on three case studies and compare it with a well-known feature location technique that uses probabilistic ranking. Our method achieves results within 83-97% of the probabilistic ranking method on the case studies without any need to create scenarios as in the existing dynamic-analysis based feature location techniques.