Assessing The Explainability of Lgbm Model For Effort Estimation on ISBSG Dataset
Merve Özdeş Demir, Bahar Gezici, Ayça Tarhan · 2023
Recently, machine learning (ML) has become very successful, which has triggered significant developments in the field of artificial intelligence (AI) and significant improvements in the predictive performance of ML algorithms. As improved performance models have emerged, the complexity of these models has increased, and this has been at the expense of human explainability of these systems, often referred to as black boxes. As a result, a new and active research field called explainable artificial intelligence (xAI) has emerged with the aim of building models with methods that are explainable and understandable to users. In this study, we focus on the explainability of the LightGBM regressor used for effort estimation by considering size attributes. We apply the SHapley Additive exPlanations (SHAP) method, a post-hoc model agnostic method, on the ISBSG dataset to shed light on the local and global explainability of the LGBM regressor. According to the results, the outputs of the SHAP method guide the user in understanding and interpreting the LGBM model on effort estimation. We concluded that there is a consistency with the results of local and global explainability.