DS-Generator: Generating Software Metrics for Subsequent Versions of a Software System

This paper introduces DS-Generator, a novel deep learning-based architecture designed to generate data associated with subsequent versions of a software system by exploiting its previous versions. DS-Generator uses a sequence prediction model to forecast future software metrics (e.g., bug counts, cyclomatic complexity) based on historical data from prior versions leading to reduced software development and testing costs, despite inherent uncertainties. The proposed architecture consists of two phases: a) data augmentation and b) next-version data generation. The effectiveness of the DS-Generator is demonstrated in terms of Mean Squared Error (MSE), Mean Absolute Error (MAE), and accuracy, we conducted experiments using eight projects from the PROMISE repository, a well-known repository in the software engineering community. The proposed architecture is compared with eight baseline methods, showing that DS-Generator achieved over 60% accuracy on 5 out of 8 software projects, with MSE and MAE ranging from 45.34 to 185.5 and 30.59 to 131.21, respectively. The results indicate that DS-Generator significantly outperforms state-of-the-art deep learning and machine learning methods, highlighting its potential to enhance efficiency in software development and testing.