Abstract
This paper provides an in-depth study and analysis of software defect prediction methods in a cloud environment and uses a deep learning approach to justify software prediction. A cost penalty term is added to the supervised part of the deep ladder network; that is, the misclassification cost of different classes is added to the model. A cost-sensitive deep ladder network-based software defect prediction model is proposed, which effectively mitigates the negative impact of the class imbalance problem on defect prediction. To address the problem of lack or insufficiency of historical data from the same project, a flow learning-based geodesic cross-project software defect prediction method is proposed. Drawing on data information from other projects, a migration learning approach was used to embed the source and target datasets into a Gaussian manifold. The kernel encapsulates the incremental changes between the differences and commonalities between the two domains. To this point, the subspace is the space of two distributional approximations formed by the source and target data transformations, with traditional in-project software defect classifiers used to predict labels. It is found that real-time defect prediction is more practical because it has a smaller amount of code to review; only individual changes need to be reviewed rather than entire files or packages while making it easier for developers to assign fixes to defects. More importantly, this paper combines deep belief network techniques with real-time defect prediction at a fine-grained level and TCA techniques to deal with data imbalance and proposes an improved deep belief network approach for real-time defect prediction, while trying to change the machine learning classifier underlying DBN for different experimental studies, and the results not only validate the effectiveness of using TCA techniques to solve the data imbalance problem but also show that the defect prediction model learned by the improved method in this paper has better prediction performance.
Highlights
With the rapid development of computer technology, software applications have expanded to all parts of people’s daily lives, creating a situation in which the economy, production, and life are fully dependent on computer software
Based on the actual defect prediction needs, defect prediction models are constructed based on the training set and selected machine learning methods; the constructed models can be used to perform defect prediction for non-training-set instances in this project; the goal of prediction is to discover whether the instances have defects or the number of defects, and this paper studies the former; that is, the instances with or without defects are predicted
The deep forest is applied to software defect prediction, and its feature vector generation method, feature transformation method, and feature enhancement method are improved to address the shortcomings in software defect prediction performance, and the software defect prediction release method based on the deep stacked forest is proposed
Summary
With the rapid development of computer technology, software applications have expanded to all parts of people’s daily lives, creating a situation in which the economy, production, and life are fully dependent on computer software. If the software development and testing stage can be combined with machine learning technology to deep search, crawl, and analyze the software’s historical defect data, to predict and count the distribution and number of defects in the software system in advance to a certain extent, it can better help the quality assurance team to understand the software quality status timely, accurately, and objectively and effectively allocate testing resources to improve software testing efficiency and save testing costs. E algorithm is based on the traditional program control flow graph combined with the system call information to automate the generation of system call transfer graph, which can effectively detect and identify malware by using the existing graph neural network model [4] When it is greater than 0.7, it shows an increasing trend, and the fine-tuned prediction model can show good prediction performance under different sparsity parameters, and the recall rate and F1-measure are better than the performance of the non-fine-tuned prediction model. Computers can automatically understand this formal language using logical inference rules
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
