Within-Project Defect Prediction Using Improved CNN Model via Extracting the Source Code Features

Abstract

Errors in software systems are inevitable. However, fixing bugs requires cost and time. If they are detected early, this problem is solved with Artificial Intelligence in Software Engineering (AISE). Therefore, software error prediction is used to discover software errors in the source code and to take into account the testing effort in the development phase. The first of the three phases presented in this work is the re-implementation of the Congs model [12] on Colab. compared the result of the original Congs model to the newly implemented model result when we found that the Congs model under Colab approximates the original results using the Simplified PROMISE Source Code (SPSC) dataset. The second is to extract deep features from source codes or ASTs as model input and proposed deep models (CNN). Third, an improved CNN model for within-project defect prediction (WPDP) by hyperparameter tuning was proposed and our results are compared to current CNN results. Our improved CNN model outperforms the improved F -measure by 29% and achieves a 93% F1-measure in our CNN model.