Feature Extraction Based Software Defect Prediction

Software defect prediction is used to detect whether the software is default,and it uses the historical version's data of current software to draw a prediction model for testing this software's error tendency.So far,many methods have been proposed,such as machine learning and statistical methods.These methods are used in software prediction and achieve good results.Based on the introduction of software defect prediction,this paper studies the feature extraction technology and makes the following work:(1)In software engineering,most of the module data is unlabeled,according to the reality,many measures have been introduced semi-supervised learning.Laplacian Eigenmaps(LE)method is an unsupervised dimension reduction method that does not take full advantage of the label data in the sample data.In this paper,the sample label is taken into account in the LE method,and a Semi-supervised Laplacian Eigenmaps method(SSLE)is proposed.This method combines labeled and unlabeled samples,and applies semi-supervised learning to LE method to effectively improve the discrimination of feature extraction.(2)LE method preserves the local information between the samples well,but the reconstructed samples may have redundant information among the samples for the resulting projection vectors are non-orthogonal.This paper proposes a Semi-supervised Holistic Orthogonal Laplacian Eigenmaps method(SSHOLE),the following is the basic idea of the method.Firstly,the objective function is constructed by using SSLE,and then orthogonal constraint is added to the objective function to get natural orthogonal projection vectors.This method effectively eliminates the redundant information between the samples and improves the prediction effect.(3)In addition,software data usually has nonlinear relationship.In order to extract the manifold structure of nonlinear data,the kernel method is used to project the samples into the high-dimensional kernel space,which makes nonlinear samples become linearly separable.In this high-dimensional linear space,the paper uses SSHOLE mentioned above,and proposes the Kernel-based Semi-supervised Holistic Orthogonal Laplacian Eigenmaps(KSSHOLE).This method can effectively improve the prediction effect of the prediction model.In this paper,the proposed methods and contrast methods are applied to NASA,AEEEM and ReLink software engineering database for testing.The experimental results validate the effectiveness of the proposed methods.Compare with the current representative software defect prediction methods,the methods proposed in this paper improve the classification performance to a certain extent.