Research On Software Defect Prediction Techniques

Along with the increasing scale and logic complexity of software system, the potential defects which are not found by testing techniques are bound to affect the quality of the system. Research on software defect prediction techniques could count the number and distribution of system defects, so as to help testers evaluate software reliability objectively, understand the software’s quality status and determine whether the software meets the delivered standard.In the thesis, based on the analysis of current software defect prediction techniques, a framework of software defect prediction based on a combination of the regression and classification method is proposed. Then this thesis does research respectively on two aspects of software complexity metric featrue selection method and defect predicion model building. The main work and innovations are as follows:1、 The existing software defect prediction framework is studied and improved. The defect prediction framework used in this thesis consists of the front, middle and end part. In the front part of the framework, the regression method is used for processing the input data sets of the software defect prediction model. In the middle part the classification method is adopted in constructing the software defect prediction model. In the end part the main work is the evaluation and comparison of different defect prediction models.2、 This thesis studies the Least Absolute Shrinkage and Selection (LASSO) method and applied the method to the area of software complexity metrics feature selection. Firstly, by studying the related concept and idea of data mining thchnique and applying them to the field of defect prediction, this thesis makes primary selection to wipe off the erroneous data. Secondly, this thesis uses the LASSO method to do the optimization selection in searching out the complexity metric attribute subsets which are most influential to defect prediction and wipe out those poor or no influence ones.3、The Least Angle Regression (LARS) algorithm and Coordinate Descent (CD) algorithm are studied and used in solving the nonlinear programming problem corresponded with LASSO method. The simulation results show that LARS algorithm is suitable for low-dimensional defect date space while CD algorithm is applicable for higher ones where the metric attribute dimension is far higher than the number of test data. Combined with the actual situation of the data set which used in this thesis, the complexity metric data subsets which selected by LASSO-LARS method are used to build defect prediction model.4、 Software defect prediction model based on the Learning Vector Quantization (LVQ) neural network optimized by the Adaptive Genetic Algorithm(AGA) is proposed. Firstly, the thesis uses the macro-search ability and global optimization performance of the AGA to solve the problem that LVQ neural network is sensitive to its initial network weights. Then, the thesis constructs defect prediction model successfully by taking advantage of the classification performance, local optimization and pattern recognition ability of the LVQ neural network. Combined with the input data subset selected by the software complexity metric feature selection method, this thesis realizes software defect prediction successfully.5、 The thesis conducts simulation experiments for the proposed software defect prediction framework by combining MATLAB platform with the relevant publicly available data set from U.S. National Aeronautics and Space Administration (NASA).The simulation results show that complexity metric feature selection method can effectively wipe out erroneous data and poor influence metrics, reduce the dimensions and at the same time obtain better defect forecast results compared with traditional feature selection methods. Also the simulation results show that the AGA-LVQ neural network based defect prediction model achieves better prediction accuracy compared with ones. Combination of two aspects of simulation experiments verifies the feasibility and effectiveness of the proposed prediction framework in this thesis.