Research On Pre-construction Of Medium And Large Project Development Pattern Based On Deep Learning

When the scale of software reaches a certain level,the number of business scenarios it contains and the complexity of associated logic between scenarios will increase exponentially with the increase of scale.Meanwhile,a medium or large project is often composed of multiple single projects.Therefore,for medium or large software,changes are accompanied by significant costs.If at the beginning of the project,through accurate software change prediction means,to build a perfect development mode with strong expansion,it can effectively reduce the number of software project change and the difficulty of change,and then improve the efficiency of the enterprise and the team,reduce the development cost.Change prediction has been a hot research direction in the field of software engineering,in the early stages of the research,some scholars put forward a project based on single project file sets in change prediction model,the model in the project of small performance is good,but for large projects,especially across projects change forecast,the accuracy and stability decline substantially.In order to solve the problems of intra-project prediction,cross-project prediction methods should be developed.However,in the practical application process,some important characteristics of projects will be greatly different,such as different programming languages,diverse technical frameworks and mutually exclusive project architectures,resulting in the final prediction effect of the model is unsatisfactory.Through in-depth study of the causes of the problems and extensive investigation of the current technical system,it is found that deep learning has strong advantages in solving the above problems.By introducing deep learning into cross-project software change prediction,this paper proposes the following two forecasting methods:DML and Innovative Beverage.That is,cross-project software change prediction method based on deep metric learning and cross-project software change prediction method based on deep ranking learning.The main research content includes the following three aspects:(1)Data preprocessing:2 closed-source projects and 5 open source projects were selected in this paper,and corresponding iteration information was extracted from the version control system,and the original data set was formed by combining the source code data.In order to solve the problem of data difference,the standardization and balance were firstly processed.Second,due to the single project model in prediction across projects performance is poorer,has an important influence factor is the single training sample data,covering scenario is less,therefore,this paper presents a project of sample collection and plan,reduce the core characteristics of the differences between individual projects,to enrich the training sample data,increase business scenario to cover,in order to improve the prediction effect,A basic experimental scheme was designed and the feasibility of project sample combination was verified by several traditional machine learning methods.(2)Cross-project software change prediction method based on deep metric learning.Software metric element is an indicator to measure the characteristics of document samples in software projects from the perspective of engineering.Therefore,how to select appropriate metric element has a direct impact on the software change prediction effect.By introducing a measure learning and deep learning the way of combining study depth measurement method is proposed in this chapter,using the method to calculate each file sample characteristic value,and the markov-distance measurement in the new mapping in the feature space,increase the distance between heterogeneous samples,similar to narrow the distance between the samples,and make the widening gap between all kinds of sample set and the corresponding threshold,Until you get the best network structure for accurate prediction of software changes across projects.(3)Change prediction method based on deep ranking learning.Although the change prediction method based on deep measure learning effectively improves the accuracy of cross-project software change prediction,its model training performance is too low and timeconsuming,which is difficult to meet the application in actual production.In order to solve the above problems,a cross-project change prediction method based on deep ranking learning is introduced.A sorting model is built to calculate the sample change probability,and the calculated results are sorted in descending order to obtain the ordered file list,so as to realize the efficient prediction of software change.Through the study of the above three aspects,the following conclusions are drawn.In the intra-project prediction,the combined training method of multiple items can effectively solve the problems of poor applicability and low accuracy of the single item training model.The experimental results show that the comprehensive index of the combined training model is improved by 12.7%.In terms of machine learning selection,the model based on deep metric learning training is better than the traditional machine learning method in change prediction.It is found in the experiment that the comprehensive index of DML increases by 6.98%yearon-year.In terms of performance,change prediction methods based on deep ranking learning are much better than deep metric learning.At the same time,its forecasting performance is also better.Innovative Beverage improved its overall indicators by 16.55%compared with deep metric learning.