Explainability of Machine Learning Using Shapley Additive exPlanations (SHAP): CatBoost, XGBoost and LightGBM for Total Dissolved Gas Prediction

Recently, explainabilityExplainability and interpretabilityInterpretability of machine learningMachine learning (ML) has been the subject of debate, and improving our understanding of ML response is becoming a challenge as various and significant factors are causing the success or the failure of the ML algorithms in solving a particular task. Undoubtedly, several techniques exist for ML explainabilityExplainability, however, the (SHAP)SHapley Additive exPlanations (SHAP) based on the game theory has received considerable research attention, and great number of application are available in the literature. In the present chapter, we present an application of various ML algorithms for predicting a real time data series: the total dissolved gas generated in high dams reservoirs, and then we tried to provide and explainabilityExplainability of the ML output using the SHAPSHapley Additive exPlanations (SHAP) technique. We select: (i) the categorical boosting machineCatBoost (CatBoost), (ii) the XGBoostXGBoost, and (iii) the LightGBMLightGBM boosting regressionRegression models for modelling TDG based on various features measured in situ. Furthermore, we realize local versus global interpretabilityInterpretability for identifying the importance of each feature, and we can conclude that the SHAPSHapley Additive exPlanations (SHAP) have helped in answering two important question: (i) making the ML interpretable by providing ranking of importance on each feature, and (ii) helped in providing a clear and concise advice about the features selection. The ML models were developed using the Outflow discharge (Q), the spill from dam (SD), water temperature (Tw), the barometric pressure (BP), and the water elevation (H). The LightGBMLightGBM was found to be the most accurate model for all stations exhibiting RMSE, MAE, R, and Nash–Sutcliffe efficiency (NSE) of approximately 1.245, 0.804, 0.983, and 0.804, respectively. Furthermore, according the results obtained using the SHAP algorithm; the Tw, Q and SD were found to be the most influencing variables on TDG concentration, however, the ranking of the input variables vary from one model to another.