Proceedings of the 6th International Conference on Predictive Models in Software Engineering. :12.
Background: Automated diagnosis of software defects can drastically increase debugging efficiency, improving reliability and time-to-market. Current, low-cost, automatic fault diagnosis techniques, such as spectrum-based fault localization (SFL), merely use information on whether a component is involved in a passed/failed run or not. However, these approaches ignore information on component execution frequency, which can improve the accuracy of the diagnostic process. Aim: In this paper, we study the impact of exploiting component execution frequency on the diagnostic quality. Method: We present a reasoning-based SFL approach, dubbed Zoltar-C, that exploits not only component involvement but also their frequency, using an approximate, Bayesian approach to compute the probabilities of the diagnostic candidates. Zoltar-C is evaluated and compared to other well-known, low-cost techniques (such as Tarantula) using a set of programs available from the Software Infrastructure Repository. Results: Results show that, although theoretically Zoltar-C can be of added value, exploiting component frequency does not improve diagnostic accuracy on average. Conclusions: The major reason for this unexpected result is the highly biased sample of passing and failing tests provided with the programs under analysis. In particular, the ratio between passing and failing runs, which has a major impact on the probability computations, does not correspond to the false negative (failure) rates associated with the actually injected faults.
