Power of LSTM and SHAP in the Use Case Point Approach for Software Effort and Cost Estimation

In this paper, we explore the effectiveness of the most popular regression models within Machine Learning (ML), such as XGBoost and MLP (Multilayer Percpetron) Regressor, and compare them with different types of Recurrent Neural Networks: LSTM (Long-Short-Term-Memory) and GRU (Gated Recurrent Unit). These models are later employed to conduct an experimental research in predicting the Real Effort of a software project using the Use Case Points (UCP) approach. Additionally, we applied a data augmentation technique to artificially increase the number of instances in the UCP Benchmark Mendeley dataset. The results have demonstrated that LSTM outperforms all the models, even though GRU is computationally more efficient and easier to train. Furthermore, we examined the SHAP (SHapley Additive exPlanations) feature importance method to improve the interpretability of the best-performing model.