Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

HyDEEP: Transparent combination of measurement and expert data for defect prediction

: Kläs, Michael
: Liggesmeyer, P.; Broy, M.; Schneider, K.


Stuttgart: Fraunhofer Verlag, 2017, XIV, 273 S.
Zugl.: Kaiserslautern, TU, Diss., 2016
PhD Theses in Experimental Software Engineering, 57
ISBN: 978-3-8396-1149-4
Fraunhofer IESE ()
systems analysis & design; industrial applications of scientific research & technological innovation; software engineering; software quality assurance; defect prediction; Softwaretechnik; Softwarequalitätssicherung; Fehlervorhersage; researchers and practitioners in software engineering; quality managers; project managers; software testers

In order to provide software products of requested quality on time and in budget, it is essential to apply analytical quality assurance (QA) activities such as reviews and tests. Planning such activities in a systematic way and controlling the effectiveness of their execution are challenging tasks for companies. The optimal selection or adjustment of a QA technique to a concrete project setting depends on knowledge about the number of defects that are contained in the artifact that will undergo QA and how effective the applied QA activity will be in finding the contained defects. Moreover, the person controlling the QA activity should know how many defects are expected to be found by the activity in order to judge whether the activity has reduced the quality risk (i.e., number of defects) as successful as expected.
In consequence, appropriately planning and controlling QA activities depends on information regarding the effectiveness of the applied activities, the defect content of the checked artifacts, and the number of defects expected to be found. However, such data is commonly missing in practical settings because these numbers strongly depend on the concrete context. Approaches to predict these numbers are rarely used in practices because they (1) require measurement data not available in most companies, (2) are too limited in their applicability, or (3) provide estimates with insufficient accuracy or trustworthiness.
This thesis therefore introduces the HyDEEP approach, which is an hybrid approach that (1) requires only commonly available measurement data and combines them with contextspecific expert knowledge in a transparent way to
(2) support the planning and controlling of static as well as dynamic QA activities. (3) Inspired by existing hybrid approaches for effort estimation, underlying principles were adapted for defect prediction and complimented with Bootstrapping to provide improved prediction accuracy and more realistic estimates of remaining uncertainty.
A study making use of legacy data of more than 60 projects indicates that the proposed Bootstrapping component can improve the realism of uncertainty estimates of hybrid models significantly. Moreover, the applicability and usefulness of the approach were evaluated in two industrial case studies. The studies showed that the approach is applicable in practical settings for different types of QA activities and requires only limited involvement of domain experts. The built prediction models were validated with contextspecific legacy data using crossvalidation, which revealed that they provide improved prediction accuracy when compared to prediction models based solely on present data.