-
This study and implementation is adapted from the study Bug Localization with Combination of Deep Learning and Information Retrieval.
-
Prepared by: Emre Dogan & Hamdi Alperen Cetin
- For our implementation, the dataset of Eclipse UI Platform is used.
- The dataset of source files is created from the original repository.
- The bug dataset can be accessed from here.
-
Before implementing this model, we prepared a survey on the machine learning applications on software bug localization area. You can find the survey from here.
-
In previous studies, a cosine similartiy based information retrieval model ,rVSM, has been used and resulted with good top-k accuracy results. In our case, rVSM approach is combined with some other metadata and fed to a deep neural network to conclude withg a relevancy score between a bug report and a source code file. This final relevany scores between all bug reports and source files are kept and top-k accuracy results for k=1,5,10,20 are calculated. In the original study, top-20 accuracy is found to be about 85% where our implementation achieves a 79% top-20 accuracy.
-
The top-k accruacy results for different k values from the original study & our study can be seen observed from the figures below.
| Original Study | Our Implementation |
|---|---|
 |
 |
More details regarding to the implementation and results can be found in the technical report.