Learning With Unlabeled Data Based Research On Software Quality Assurance

With the increasing integration level of software into daily life,people get higher requirements for software quality.In order to improve the software quality in the perspective of defects as the object and developers as the subject,thorough research and analysis on software defect detection and effort estimation have been conducted respectively,and new methods based on learning with unlabeled data have been utilized to tackle the issue of scarcity of valid labeled data,with some major results obtained and described as follows.First of all,as to the issue of biased label in software defect detection,by assuming these modules as unlabeled data,the positive and unlabeled learning framework is utilized to identify hidden defective modules and a new technique based on kernel density estimation is proposed to extract reliable negative instances.The proposed method is capable of identifying hidden defective modules from the originally labeled as defect-free ones effectively.Furthermore,a defect detection model in a semi-supervised manner is built by using the remaining unlabeled data to further improve the prediction performance to a large extent.Secondly,to tackle the effort estimation issue in software development with extremely small data sets,a novel technique based on the twice learning framework is proposed to generate a large amount of unlabeled virtual examples,and combine the models with strong generalization ability and high comprehensibility respectively to build the ultimate prediction model.As a result,the proposed model is able to achieve better performance as well as disclosing the key factors within effort estimation effectively.