A Research On Fine-grained Software Defect Prediction Method

As software has become larger and more complex,software quality assurance become increasingly important.In software development process,software testing is indispensable for locating and repairing software defect to improve software quality.Software defect prediction predicts defective modules effectively,which helps software testing engineers discover more defect by prioritizing testing resource for defect-prone modules.The majority of studies performed defect prediction at subsystem level,pack-age level,file level or class level.Conducting defect prediction at finer granularity can narrow down the location of defects,which reduces resource to test modules predicted to defect-prone.Thus fine-grained defect prediction is capable to improve testing ef-fort allocation significantly and save manual inspection effort.Previous research has confirmed that performing fine-grained defect prediction makes prediction model has high cost-effectiveness.Performing fine-grained defect prediction has become a cur-rent trend in the area of software defect prediction.This thesis studies the problems in fine-grained defect prediction.According to the static hierarchy structure of software,we define horizontal fine-granularity as class and method level,vertical fine-granularity as execution path level,achieving the following innovations:1.When performing defect prediction at horizontal fine-granularity,as the granu-larity of prediction becomes finer,the number of modules to be predicted is larger and the proportion of defective modules is less.In this situation the collection of labeled data become more expensive and difficult,even no available labeled data sometimes.To solve this problem,in the thesis we proposed a fine-grained defect prediction ap-proach based on anomaly detection by regrading defect as anomaly,which can predict defect without labeld data.The experimental results indicate that our defect prediction method achieves better performance compared with defect prediction model based on supervised learning or semi-supervised learning when labeled rate is no more than 5%.2.Path covering is a usually used method for white box testing.If we can pre-dict defect at vertical fine-granularity,i.e.at execution path level effectively,testing resource for path covering would be saved significantly.However,there is no research performs defect prediction at execution path level.To solve this problem,in the the-sis we proposed a novel fine-grained defect prediction model,i.e.MI-CNN,based on multi-instance learning.Experimental results show that our model ahcieved the best performance of defect prediction at method level.Furthermore,MI-CNN can perform defect prediction at execution path level effectively.