Authors:
Arpita Dutta
1
and
Rajib Mall
2
Affiliations:
1
School of Computing, National University of Singapore, Computing Dr, Singapore
;
2
Dept. of Computer Science and Engineering, Indian Institute of Technology, Kharagpur, India
Keyword(s):
Software Fault Localization, Debugging, Ensemble Classifier, Program Analysis.
Abstract:
Fault localization (FL) is one of the most difficult and tedious task during software debugging. It has been reported in literature that different FL techniques show superior results under different circumstances. No reported technique always outperforms all existing FL techniques for each type of bug. On the other hand, it has been reported that ensemble classifiers combine different learning methods to obtain better predictive performance that may not be obtained from any of the constituent learning algorithms alone. This has motivated us to use an ensemble classifier for effective fault localization. We focus on three different families of fault localization techniques, viz., spectrum based (SBFL), mutation based (MBFL), and neural-network based (NNBFL) to achieve this. In total, we have considered eleven representative methods from these three families of FL techniques. Our underlying model is simple and intuitive as it is based only on the statement coverage data and test execut
ion results. Our proposed ensemble classifier based FL (EBFL) method classifies the statements into two different classes viz., Suspicious and Non-Suspicious set of statements. This helps to reduce the search space significantly. Our experimental results show that our proposed EBFL technique requires, on an average, 58% of less code examination as compare to the other contemporary FL techniques, viz., Tarantula, DStar, CNN, DNN etc.
(More)