Abstract:
Empirical studies on software defect prediction
models have come up with various predictors. In this study we
examined variable regularized factors in conjunction with
Logistic regression. Our work was built on eight public NASA
datasets commonly used in this field. We used one of the datasets
for our learning classification out of which we selected the
regularization factor with the best predictor model; we then
used the same regularization factor to classify the other seven
datasets. Our proposed algorithm Variant Variable Regularized
Logistic Regression (VVRLR) and modified VVRLR; were then
used in the following metrics to measure the effectiveness of our
predictor model: accuracy, precision, recall and F-Measure for
each dataset. We measured above metrics using three Weka
models, namely: BayesianLogisticRegression, NaiveBayes and
Simple Logistic and then compared these results with VVRLR.
VRLR and modified VVRLR outperformed the weka
algorithms per our metric measurements. The VVRLR
produced the best accuracy of 100.00%, and an average
accuracy of 91.65 %; we had an individual highest precision of
100.00%, highest individual recall of 100.00% and F-measure of
100.00% as the overall best with an average value of 76.41%
was recorded by VVRLR for some datasets used in our
experiments. Our proposed modified VVRLR and variant
VVRLR algorithms for F-measures outperformed the three
weka algorithms.
