Defect Prediction using Exception Handling Method Call Structures

Abstract

The main purpose of exception handling mechanisms is to improve software robustness by handling exceptions when they occur. However, empirical evidence indicates that improper implementation of exception handling code can be a source of faults in software systems. There is still limited empirical knowledge about the relationship between exception handling code and defects. In this dissertation, we present three case studies investigating defect densities of exception handling code. The results show that in every system under study, the defect density of exception handling code was significantly higher than the defect density of overall source code and normal code. The ability to predict the location of faults can assist in directing quality enhancement efforts to modules that are likely to have faults. This information can be used to guide test plans, narrow the test space, and improve software quality. We hypothesize that complicated exception handling structure is a predictive factor that is associated with defects. To the best of our knowledge, no study has addressed the relationship between the attributes of exception handling method call structures and defect occurrence, nor has prior work addressed fault prediction. We extract exception-based software metrics from the structural attributes of exception handling call graphs. To find out whether there are patterns of relationship between exception-based software metrics and fault-proneness, we propose a defect prediction model using exception handling call structures. We apply the J48 algorithm, which is the Java implementation of the C4.5 algorithm, to build exception defect prediction models. In two out of three systems under study, the results reveal that there are logical patterns of relationships between most class level exception metrics and fault-proneness. The accuracy of our prediction models is comparable to the results of defect prediction model studies in the literature. It was observed that our approach has somewhat worse predictive accuracy when a system has low average defects per class.