CROSS PROJECT DEFECT PREDICTION

Abstract

Cross-project defect prediction (CPDP) recently gained considerable attention. Many studies have provided the success of cross-project defect prediction (CPDP) to predict defects. But, most of the datasets share the same limitations: as the metrics (independent variable) hardly follow a normal distribution. They are mostly skewed data. Hence, these metrics has to be transformed before the training and predicting the defect. Various transformations have been studied in the area of cross-project defect prediction like rank transformation, log transformation and Box-Cox transformation. The yeojohnson transformations (extended versions of the box-cox transformation) have not been used in the defect prediction area. Since, the metric values contain Zero as a value so most of the transformation did not work for Zero value, so we need to do some precomputation in data. But yeo-johnson transformation can be used for zero values as well as negative value. This study investigates the effectiveness of yeo-johnson transformation on cross-project defect prediction. we have conducted our experiment on publicly available Promise data sets. Comparing logistic regression model built using with and without yeo-johnson transformation, transformed approach gives better result than raw data in 53% cases and 70% project achieved better result using this transformation. Further, we also investigate which classifier is well suited for this transformation, which is statistically tested by Friedman's test. Naïve bayes outperforms better among all classifiers.