Defect prediction is an important task for preserving software quality.
A lot of previous research has analyzed source code to predict defects; however,
it contains a problem that is its prediction grain is too coarse and its feedback is too late
for software developers.
To achieve a more fine-grained prediction and an earlier feedback, several approaches that analyzes source code changes has been reported.
Those approaches have applied various machine learning techniques and
deep learning techniques to change metrics, such as the number of lines added,
modified files, and modified directories.
In this paper, we propose a novel approach for defect prediction called Word-Convolutional Neural Network (W-CNN), which applies CNN to the modified source code itself.
Our evaluation results show that the proposed approach can improve the effectiveness of defect prediction with a small overhead on the prediction time.
- Masanari Kondo, Keita Mori, Osamu Mizuno, Eun-Hye Choi, "Just-In-Time Defect Prediction Applying Deep Learning to Source Code Changes," 情報処理学会論文誌, 59(4), pp. 1250-1261, April 2018.
- Masanari Kondo, Keita Mori, Osamu Mizuno, Eun-Hye Choi, "深層学習による不具合混入コミットの予測と評価," ソフトウェアエンジニアリングシンポジウム2017論文集 (SES2017) , pp. 35-44, August 2017. (東京都)
- Keita Mori, "An Application of Deep Learning Based Classifier to Defect Prediction," Master thesis, 京都工芸繊維大学 大学院工芸科学研究科, 2017.
- Masanari Kondo, Keita Mori, Osamu Mizuno, Eun-Hye Choi, "最優秀論文賞: 深層学習による不具合混入コミットの予測と評価," ソフトウェアエンジニアリングシンポジウム2017 (SES2017), August 2017.