Software Defect Prediction using Adaptive Neural Networks

We present a system which gives prior idea about the defective module. The task is accomplished using Adaptive Resonance Neural Network (ARNN), a special case of unsupervised learning. A vigilance parameter (θ) in ARNN defines the stopping criterion and hence helps in manipulating the accuracy of the trained network. To demonstrate the usefulness of ARNN, we used dataset from promisedata.org. This dataset contains 121 modules out of which 112 are not defected and 9 are defected. In this dataset modules are termed as defected on the basis of three measures that are LOC, HALSTEAD, MCCABE measures that have been normalized in the range of 0-1.We see that at θ=0.1858 the network has maximum Recall (i.e. true negative rate) is 100% and average Precision=54%.In case of ART n/w shortfalls are seen for Accuracy as this is a subjective measure. Keyword: Resonance, Clustering, Unsupervised learning,