Fine-Grained Fault-Proneness Prediction

Software defect prediction models are built with software metrics which can be used for predicting latent defects in software systems.With these defect prediction models,we can automatically allocate testing or review effort for software quality con-trol.Up till now,a large amount of research effort have been devoted to developing better defect prediction models.However,there still exist some problems that need to be resolved.First,the performance of defect prediction models is not well enough since they are built with limited code and process metrics.This will prevent the wide use of defect prediction models in practice.Second,most defect prediction models are at coarse-granularity(e.g.,package-level,file-level,or class-level).The prediction at coarse-granularity is not practical enough since we cannot narrow down the area of defect location.In order to solve these problems,we empirically investigate three kinds of fine-grained defect prediction models.We aim to improve the prediction performance and make defect prediction models more practical.In this thesis,our investigated defect prediction models are at function-level,change-level,and line-level.Compared with coarser-grained defect prediction models,fine-grained models can narrow down the defect areas which will provide more useful information.Our contributions are sum-marized as follows:1)At function-level defect prediction,we empirically investigate the usefulness of slice-based cohesion metrics in effort-aware function-level defect prediction.Our experimental results from several open-source software systems show that,when used together with traditional code and process metrics,slice-based cohesion metrics can effectively improve the effort-aware defect prediction performance at function-level.2)At change-level defect prediction,we empirically investigate simple unsuper-vised models in effort-aware defect-inducing change prediction in both 10 times ten-fold cross-validation,time-wise cross-validation,and across-project prediction.Based on data sets from six open-source software systems,our results show that many simple unsupervised models perform similarly to or significantly better than the state-of-the-art supervised effort-aware defect-inducing change prediction models.3)At line-level defect prediction,we empirically investigate a line-level defect prediction model based on the number of words in source code.Based on several open-source systems,our experimental results show that this simple model performs better than the state-of-the-art N-grams based line-level defect prediction model.