Research On Prediction Method Based On Manifold Learning For Object Oriented Software Defect

With the scale of projects development have increasingly large, software developmenttechnology gradually become object-oriented development, in order to better describe thecharacteristics of object-oriented software, software metric model need to use more attributesto measure software which lead to the dimensions of data is higher and higher. However,with the increasing of the dimension of data, effective predict software defects becomesmore and more difficult, resulting in a “dimension disaster’’ problem. Therefore, in order topredict the existence of defects in the software more accurately so as to improve the qualityof the software, software metrics for high-dimensional data reduces dimensionality isnecessary. Manifold learning method is an important means for dealing withhigh-dimensional data, it can be found hidden in the structure of software metrics ofhigh-dimensional data. This paper mainly studies how to use the manifold learning is appliedto object-oriented software defect model, research contents include the following aspects:1、 Analysis and Comparing the effect of the software defect prediction method.Including SVM, NB, KNN and BP neural network. With the widely application of theobject-oriented technology, in order to describe the characteristics of object-orientedsoftware more comprehensive so need to use more metrics properties, lead to increasinglyhigh dimension of measurement data for prediction software defect, when the softwaredefect prediction data present a high dimensional feature, those prediction methods cannotachieve good prediction effect.2、For high-dimensional nature of the data on the impact of predicted results, so proposea prediction model based on manifold learning for object-oriented software defect. In thismodel, firstly use of LLE, LE, ISOMAP, PCA manifold learning algorithms extractlow-dimensional feature from object-oriented software defect data. Secondly, use traditionalsoftware defect prediction method for low-dimensional feature classification. Through twodata sets verify the validity of the proposed model, the experimental results show that thedimensionality reduction and then classify not only improves the prediction accuracy ofprediction methods, but also greatly improve the efficiency of prediction methods.3、Through the experiment on the manifold learning algorithm to estimate k nearestneighbor number and d dimensionality. Choose the optimal parameter according to theexperimental results. And compared the use of several kinds of manifold learning algorithmto extract the features on the result of prediction, can be found by the result of theexperiment, LE algorithm is utilized to extract the characteristics of the best results.