EFSPredictor: Predicting configuration bugs with ensemble feature selection

• Main Authors: XU, Bowen, David LO, XIA, Xin, SUREKA, Ashish, LI, Shanping
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2016
• Subjects: Configuration Bugs; Data Mining; Ensemble Feature Selection; Databases and Information Systems; Data Storage Systems
• Online Access: https://ink.library.smu.edu.sg/sis_research/3641; https://ink.library.smu.edu.sg/context/sis_research/article/4643/viewcontent/9644a206.pdf
• Institution: Singapore Management University

The configuration of a system determines the system behavior and wrong configuration settings can adversely impact system's availability, performance, and correctness. We refer to these wrong configuration settings as configuration bugs. The importance of configuration bugs has prompted many researchers to study it, and past studies can be grouped into three categories: detection, localization, and fixing of configuration bugs. In the work, we focus on the detection of configuration bugs, in particular, we follow the line-of-work that tries to predict if a bug report is caused by a wrong configuration setting. Automatically prediction of whether a bug is a configuration bug can help developers reduce debugging effort. We propose a novel approach named EFSPredictor which applies ensemble feature selection on the natural-language description of a bug report. It uses different feature selection approaches (e.g., ChiSquare, GainRatio and Relief) which output different ranked lists of textual features. Next, to obtain a set of representative textual features, EFSPredictor first assigns different scores to the features outputted by these feature selection approaches. Next, for each feature, EFSPredictor sums up the scores outputted by the multiple ranked lists, and outputs the top features (e.g., 25% of the total number of features) as the selected features. Finally, EFSPredictor builds a prediction model based on the selected features. We conduct experiments on 5 bug report datasets (i.e., accumulo, activemq, camel, flume, and wicket) containing a total of 3,203 bugs. The experiment results show that, on average across the 5 projects, EFSPredictor achieves an F1-score to 0.57, which improves the state-of-the-art approach proposed by Xia et al. by 14%.