Toward Extended Change Types for Analyzing Software Faults

Abstract

Abstract-This research extends an existing source code change taxonomy that was designed to analyze change coupling. The extension expands change types related to statements in order to achieve more granular data about the type of statement that is changed. The extended taxonomy is evaluated to determine if it can be applied to software fault analysis. We found that the extended change types occur consistently and with high frequency in fault fixes for Eclipse 2.0 and 3.0. Faults were then clustered according to the source code changes and analyzed. We found that the types and sizes of clusters are highly correlated, indicating some consistency in the patterns of the fault fixes. Finally, we performed an initial investigation to determine whether faults in the same cluster have similar characteristics. Our results indicate that many of the change types can be used to characterize the type of fault that has been fixed. However, some of the change types obfuscate the true nature of the fix. Ideas for improving the taxonomy based on these findings are provided.

Toward a learned project-specific fault taxonomy: application of software analytics

“…The tool and taxonomy were created for change impact analysis. We were able to extend the taxonomy using contextual information that is captured by the tool [13]. Additional changes were made to the ChangeDistiller project in order to handle problematic constructs such as the presence of anonymous classes.…”

“…The cosine similarity is used as the distance measure. A repeated bisection clustering algorithm was used with the I1 criterion function [13]. CLUTO reports the features in the vector that contribute to the internal similarity of each 1 https://bitbucket.org/bill_kidwell/tools-changedistiller/ cluster [14].…”

“…Initial manual analysis indicates that the nature of faults within a cluster is indeed similar: i.e., we have identified a fault type that is part of a larger fault taxonomy. For more details about the resulting clusters, data collection, and validation please refer to our previous publication [13].…”