Hierarchical multi-scale dense-RNN with sparse attention model for effort and cost estimation of IT projects in agile methodologies

The existing models face significant limitations, including an inability to adapt to changing project variables, potential generalization issues when using small datasets, and high computational costs for optimization. Accurately validating project cost remains a primary complexity. In this article, initially, input data is collected from online sources like Software Cost Estimation 1 (Dataset 1) and the SEERA Software Cost Estimation Dataset (Dataset 2). It fed into Hierarchical Multi-scale Dense-Recurrent Neural Network with Sparse Attention (HMDRSA) model. This HMDRSA architecture improves upon existing deep learning models by integrating a unique combination of hierarchical processing, multi-scale analysis, dense connections, and a sparse attention mechanism. This deep learning architecture is particularly well-suited for agile SCRUMBAN environments because its architecture is designed to capture complex project patterns efficiently, significantly reduce computational overhead, and improve generalization compared to traditional methods that struggle with dynamic project variables. Thus, the model aids in managing different Development Operations (DevOps) activities, helping to deliver software projects more quickly with higher quality. Finally, its performance is validated by comparing it with existing models. Experimental results on two public datasets demonstrate the model’s effectiveness, achieving low error rates with an MAE of 3.01 and 3.89, and RMSE of 10.08 and 11.42.