Research On Software Defect Prediction Based On Code Semantic And Structural Features Learning

With the rapid development of Internet technology,people can feel the arrival of information and intelligence in various aspects such as clothing,food,housing and transportation,and enjoy the convenience brought by information products.However,every coin has two sides.Behind its convenience,once the software security problems occur,there will be different levels of serious costs and sacrifices.This requires a large number of industry personnel to review software system security defects,predict the potential defects in the code,guide the testing work,reduce development costs,so as to improve software quality.In traditional software defect prediction,some metrics are designed manually or combined with existing metrics as code features,such as Mc Cabe metric,CK metric,Mood metric,etc.,and then classifier training is carried out through machine learning algorithm.However,the program information that these features can capture is very limited.If more advanced technology can be used to detect software security risks,it is of great significance to ensure the stability and convenience of social life and work.Therefore,in view of the problem that small differences in the order of local code will lead to huge differences in the global program,this paper adopts deep learning method to extract semantic features,but these fine-grained semantic features ignore the global structure information between source code files,and adopts network embedding method to automatically learn structural features.At the same time puts forward a comprehensive consideration of code semantic and structural features of software defect prediction method,the source code files generated by said learning,on the basis of semantic characteristics and structural characteristics,characteristics of the two for further learning and to strengthen,using multiple fusion methods to adopt the deep learning method to extract classification after the fusion,defects of software source code files.In this paper,6 open defect data sets provided by Promise platform were taken as experimental objects,and the effect of defect prediction method presented in this paper was compared and analyzed by using F-measure index.Experimental results show that,compared with the existing methods,the proposed method can improve the F-measure value by up to6.74%.Meanwhile,the parameters in the learning process of semantic features and structural features are analyzed to provide guidance for the optimization of the predictor.