On the Use of CTGAN Data Augmentation for Enhancing MLP-based Software Development Effort Estimation using small datasets

Abstract

Accurate effort estimation is vital for successful software project management, yet limited historical data often hinders the performance of estimation models. This study investigates the impact of CTGAN-based data augmentation on a Multilayer Perceptron (MLP) architecture for Software Development Effort Estimation (SDEE). We assess the performance of an MLP model augmented with synthetic data generated by CTGAN using five benchmark datasets, namely ISBSG, COCOMO, Albrecht, Desharnais, and China. The performance of the CTGANaugmented MLP is compared to a baseline MLP trained without augmentation using key metrics like Pred(25), MAE, MMRE and MSE. According to the experimental finds, CTGAN augmentation improves estimation accuracy substantially on most of the datasets. In this context, Pred(25) scored $\mathbf{9 8 \%}$ on ISBSG and $\mathbf{9 6 \%}$ on COCOMO. However, the performance drops of the China dataset demonstrate how the characteristics of a dataset affect the efficacy of augmentation. The research findings indicate that the use of CTGAN-based data augmentation is a viable approach for advancing the performance of effort estimation models in the presence of scarce historical data.