Explainable Artificial Intelligence Methods for Modeling Categorical Responses

Artificial Intelligence (AI) and Machine Learning (ML) are technologies that are not only among the top fields of research, but their areas of applicability are also rapidly increasing. For instance, complex non-linear or non-parametric ML models are already being applied to support key decisions in sectors like health care, criminal justice, or finance. However, while using ML or AI in critical sectors can be viewed as favorable due to their generally high predictive performance and power, the main drawback for human users is that drawinginference, detecting bias or retracing the basis for the different models’ decisions is often hard, if not impossible. The field of Explainable AI addresses exactly this prevalent lack of interpretability of various complex machine learning models – and multiple proposals on how to solve this issue have already been made. In this work we discuss three different methods of Explainable AI, which allow for an interpretation of complex models in a way that is accessible to humans. Moreover, we compare the methods in terms of performanceand applicability on a use-case scenario, namely to predict the electricity imbalance price for Austria, using data collected from the Transparency Platform, which is operated by the European Network of Transmission System Operators for Electricity (ENTSO-E). Specifically, the considered methods are Local Interpretable Model-agnostic Explanations (LIME), Shapley values for model explainability, and SHapley Additive exPlanations (SHAP), for all of which the focus is placed on their application in the context of classification tasks. Re-garding the classification of the electricity imbalance price, we analyze the time-dependent electricity market data from Austria and explore multiple modeling strategies. While all three methods of Explainable AI discussed in this work provide model-agnostic explanation procedures for individual observations, only the model-specific Tree SHAP algorithm from the SHAP framework offers efficient implementations that enable us to fully explain all predictions of tree-based models. The final model for the classification of the electricity imbalance price is based on an online learning approach utilizing boosted tree ensemblemodels, for which we employ the methods of Explainable AI to interpret the model. Furthermore, tree-based methods do not only yield the best results for the currently available data, but also possess the highest potential for improvement should the data quality or availability improve in the future.