The approach

You have decided to use local model interpretation to explain why each bar is rated as it is. To that end, you will prepare the dataset and then train classification models to predict if chocolate-bar ratings are above or equal to Highly Recommended, because the client would like all their bars to fall above this threshold. You will need to train two models: one for tabular data, and another NLP one for the words used to describe the chocolate bars. We will employ support vector machines (SVMs) and Light Gradient Boosting Machine (LightGBM), respectively, for these tasks. If you haven't used these black-box models, no worries—we will briefly explain them. Once you train the models, then comes the fun part: leverage two local model-agnostic interpretation methods to understand what makes a specific chocolate bar less than Highly Recommended or not. These methods are SHAP and LIME, which when combined will provide a richer explanation to convey...