Synergetic interaction between fault localisation and defect prediction

Abstract

Identifying the root cause of a program failure (i.e., fault) is crucial for maintaining software quality. Fault localisation intends to identify faults after they are revealed through a program failure, whereas defect prediction aims to predict yet-to-happen faults. Although they both aim to identify faults in code, just with different timing, fault localisation and defect prediction have been mainly studied as separate research topics, and thereby the synergy between them remains largely under-explored. This thesis argues that the synergetic interactions between fault localisation and defect prediction can enhance both techniques, as they share the common goal of identifying faulty code. To validate this claim, we first investigate whether defect prediction can improve fault localisation and vice-versa. The empirical results show that fault localisation and defect prediction do enhance each other: leveraging code and change features widely studied in defect prediction allows at least 22% more fault to be precisely localised, whereas using the code suspiciousness computed from past fault localisation can improve both the accuracy and the actionability of defect prediction. Reducing the debugging cost is crucial for industrial software systems, as it directly affects the profit of a company. Nevertheless, existing fault localisation techniques expected to reduce this cost have been rarely evaluated with industrial software systems. This thesis also includes the industrial scale case study of automated fault localisation techniques. The results of the case study show that existing fault localisation techniques can assist the debugging process of industrial projects by finding faults the original approach in the projects was unable to locate.