Impact of using information gain in software defect prediction models

Abstract

Presence or absence of defective modules in software is an indicator of quality of the software. Every company aspires to deliver good quality software with minimum number of defective modules. To achieve this goal, defect prediction models are used in different phases of software lifecycle. These models have to deal with a large number software metrics (as input parameters to the models). These metrics have correlation issues that affect a model's performance. Also, in some cases using all the metrics negatively impacts the models' performances. In order to reduce size of input space and resolve the possible issues of correlation in input data, models reported in literature use Principal Component Analysis (PCA) and Information Gain (IG) based dimension reduction. PCA reduces the dimensions but keeps the representation of all the input variables intact. Use of PCA is not suitable where representation of all the metrics is declining a model's performance. To handle such situations, this paper advocates use of Information Gain (IG) based technique to reduce size of input space by dropping the irrelevant metrics. Afterwards, only relevant metrics are used to develop a prediction model. This paper compares the PCA and IC based techniques to develop classification tree and fuzzy inferencing system based models. In order to study the impact of using IG, percentage improvement in Recall, Accuracy and Misclassification Rate have been calculated for the aforementioned models. The results show that use of IG improves the models' performances more often than PCA does. © 2014 Springer International Publishing Switzerland.