Modern machine learning methods lack of interpretability, it is often unclear why a certain input leads to a certain output. Over the last years SHAP has become a popular method to address this issue. The models that we want to explain are introduced, GLMs and LightGBMs. We explain the theoretical concepts behind SHAP and how it is related to game theory, then we find closed form solutions for the models that are currently used at Allianz. The issue of computational complexity is also discussed, SHAP values can be obtained by regression. We give an overview about different methods of removing features and argue, why the currently implemented way of SHAP is the right one for us. The SHAP package is presented and its specific plots are introduced, we provide some hints about working with the package. We propose a new method to detect interactions based on SHAP values. A visualization tool to interactively work with SHAP values is presented. Finally, we summarize the results of our case study, where we explained three models with SHAP and compared the results; based on the SHAP values of the LightGBM models we tried to improve the existing Emblem model.     «

Modern machine learning methods lack of interpretability, it is often unclear why a certain input leads to a certain output. Over the last years SHAP has become a popular method to address this issue. The models that we want to explain are introduced, GLMs and LightGBMs. We explain the theoretical concepts behind SHAP and how it is related to game theory, then we find closed form solutions for the models that are currently used at Allianz. The issue of computational complexity...     »