Research And Application Of Feature Description In Software Defect Prediction

As software products widely applied in industrial manufacturing, finance,transportation and military, as software systems grow in size and complexity, how tomaintain the quality and reliability of software products has become the important anddifficult problem in research of software engineering. Software defect is one of theimportant factors influencing software reliability. Predicting software defect duringsoftware test could estimate the distribution of software defect at a lower cost andconcentrate the major human and financial resources on the test of module with defect,which is very valuable for ensuring software quality and reliability.Classification technique is the main software defect prediction method. Featuredescription and classifier choice are the most two important factors in model ofclassification and prediction. Based on the characteristics of software defect data, weexplore novel software module description methods for software defect prediction.Taking NASA data set for example, our aim is to realize improvement in performanceof software defect prediction. The main work is as follows:①Aiming to settle the problems of multiple types of software module features,low accuracy of prediction model and high computational complexity of algorithm infeature space, dissimilarity space based software defect prediction model is proposed.Firstly, a representation set is generated from training data. Then, we select appropriatedissimilarity measure, and translate feature space description based software modulesinto dissimilarity space based ones. Finally, classifier conducted on dissimilarity spaceis used to complete software defect prediction. Experimental results on NASA data setsshow that dissimilarity description could guarantee the accuracy of software defectprediction and effectively reduce the computational complexity. In CM1, KC1, KC2andPC1, the computation time is improved by1%to67%.②From the perspective of increasing the classification effectiveness of eachsoftware metric, we propose classifying feature description for software module. Firstly,independent classifier is constructed on each software metric. Then the classifyingresults in each feature are used to represent each feature of software module. Finally,prediction model is conducted based on classifying feature description. To obtain theclassifying feature description of software modules, we propose two different featureclassifier algorithms, which are based on mean criterion and minimum error rate criterion. Experiment results demonstrate classifying feature description significantlyimprove performance of software defect prediction. The best accuracies in the four datasets are improved from65.61%,71.79%,69.67%and65.01to71.29%,75.99%,78.05%and73.96%, respectively.③Considering the facts that various types of software metrics are estimateddifferently and they usually have different distributions. Aggregation feature descriptionof software module is proposed. To overcome the drawbacks of high computationalcomplexity, low anti-noise performance and ignorance of weights of various types offeature in SVM based aggregation algorithm, we construct an aggregation frameworkbased on LS-SVM classifier and Boosting algorithm. Firstly, LS-SVM classifier isconducted on each type of feature to obtain the aggregation feature. Then, Boostingalgorithm is used to calculate weight of aggregation feature. Finally, prediction model isconducted on weighted aggregation feature description software module. Experimentalresults show that aggregation feature description performs better than regular featuredescription. The best accuracies in the four data sets are improved to74.04%,75.05%,77.22%and75.21%, respectively.