Improving Quality of Code Review Datasets – Token-Based Feature Extraction Method

Paper in proceeding, 2021

Machine learning is used increasingly frequent in software engineering to automate tasks and improve the speed and quality of software products. One of the areas where machine learning starts to be used is the analysis of software code. The goal of this paper is to evaluate a new method for creating machine learning feature vectors, based on the content of a line of code. We designed a new feature extraction algorithm and evaluated it in an industrial case study. Our results show that using the new feature extraction technique improves the overall performance in terms of MCC (Matthews Correlation Coefficient) by 0.39 – from 0.31 to 0.70, while reducing the precision by 0.05. The implications of this is that we can improve overall prediction accuracy for both true positives and true negatives significantly. This increases the trust in the predictions by the practitioners and contributes to its deeper adoption in practice.