Quantifying Explainability in Outcome-Oriented Predictive Process Monitoring

Abstract The growing interest in applying machine and deep learning algorithms an Outcome-Oriented Predictive Process Monitoring (OOPPM) context has recently fuelled a shift to use models from the explainable artificial intelligence (XAI) paradigm, field of study focused on creating explainability techniques top AI order legitimize predictions made. Nonetheless, most classification are evaluated primarily performance level, where XAI requires striking balance between either simple (e.g. linear regression) or using complex inference structures neural networks) with post-processing calculate feature importance. In this paper, comprehensive overview predictive varying intrinsic complexity measured based model-agnostic quantitative evaluation metrics. To end, is designed as symbiosis interpretability faithfulness thereby allowing compare inherently created explanations decision tree rules) post-hoc Shapley values) models. Moreover, two improved versions logistic regression model capable capturing non-linear interactions both generating their own proposed OOPPM context. These benchmarked common state-of-the-art explanation explainability-performance space.