Automatic defect categorization based on fault triggering conditions

Due to the complexity of software systems, defects are inevitable. Understanding the types of defects could help developers to adopt measures in current and future software releases. In practice, developers often categorize defects into various types. One common categorization is based on fault triggers of defects. Fault trigger is a set of conditions which activate a defect (i.e., Fault) and propagate the defect into a failure. In general, there are two types of defect based fault triggering conditions, Bohrbug and Mandelbug. Bohrbug refers to a bug which can be easily isolated, and its activation and error propagation is simple. Mandelbug refers to a bug whose activation and/or error propagation is complex (e.g., A time lag between the fault activation and the failure occurrence). With these category labels, developers can better perform post-mortem analysis to identify common characteristic of the defects, and design specific fault-tolerance mechanisms. However, in most software systems, these category labels are often unavailable. To address this problem, in this paper, we propose a text mining solution which categorize defects into fault trigger categories by analyzing the natural-language description of bug reports. A previous study shows that Mandelbug is more complex and needs more time to be fixed. Thus, to better identify Mandelbugs, we propose a novel Fuzzy Set based Feature Selection algorithm named USES, which selects the features (i.e., Terms) which have high ability to distinguish Mandelbugs from Bohrbugs. USES first caches a set of terms based on their fuzzy affinity scores to Bohrbug or Mandelbug. Next, it iterates many times, and in each iteration, it selects a subset of terms, and builds a classifier on these terms. USES selects the classifier and the terms which could achieve the best performance on a training data. We evaluate our solution on 4 datasets including Linux, Mysql, Apache HTTPD, and AXIS containing a total of 809 bug reports. We show that USES with naive Bayes multinomial achieves the best performance, it achieves Mandelbug F-measure scores of 0.298-0.615. We also compare USES with other baseline approaches. The results show that USES on average improves Mandelbug F-measure scores of the best performing baseline by 12.3%.