Improving Software Defect Assignment Accuracy with the LSTM and Rule Engine Model
Abstract
After a software defect is reported with a title and a text description, a competent developer needs to be assigned to fix it. The accuracy of this assignment has big impact on the quality of the resulting software, and the speed of the debugging process. Traditionally this software defect assignment process is conducted by product managers based on his/her knowledge of the software and the developers, which is not very scalable. In the recent years, this defect assignment problem has been formulated as a problem of (1) feature extraction from the defect title and description, and (2) classification of the resulting feature sets to the developers. Machine learning has been used to automate this software defect assignment problem. The research improves the existing approaches in automatic defect assignment by (1) improving the feature extraction by NLP and Vector for Words technology, (2) introducing rule-based engine aka expert system to better character the strength of each developer, instead of the traditional characterizing a developer only by the descriptions of the bugs he/she has resolved; (3) combining the two layers model of our model (Layer 1, NLP and Vector for Words and Layer 2, Long Short-Term Memory and Rule-based Engine). The optimal results are achieved on the CHROME dataset based on our new model of Long Short-Term Memory(LSTM) with Rule-based Engine in comparison with the traditional ML model - Naïve Bayes model. The proposed neural network model extracts text features on its own, considering not only the word order messages that the word bag model ignores, but also the grammatical and semantic characteristics of the text. Rule-based Engine has absorbed developers’ history data, and activity information. The structure of these two layers network model with Rule-based Engine is relatively simple, i.e. the model is parallel structure, ideal for parallel computing, plus a dedicated hardware processing accelerator GPU makes the model not only high accuracy, but also faster. The new approach that we introduced in the research shows that it has better accuracy than traditional Naïve Bayes model and pure LSTM model. The new model can expand and migrate the system to generic bug assignment problems. The model is expandable and migratable.
Subject Area
Computer science|Computer Engineering|Mathematics
Recommended Citation
Zhu, Robert, "Improving Software Defect Assignment Accuracy with the LSTM and Rule Engine Model" (2019). ETD Collection for Pace University. AAI27543321.
https://digitalcommons.pace.edu/dissertations/AAI27543321
Remote User: Click Here to Login (must have Pace University remote login ID and password. Once logged in, click on the View More link above)