Explainability in knowledge discovery from data streams
Szymon Bobek, Grzegorz Jacek Nalepa · 2019
Interpretability of machine learning algorithms become an emerging research area over last decade. Most powerful algorithms such as deep neural networks, gradient boosting trees, deep reinforcement learning and other do not offer human-understandable justification of their decisions. However, such explanations are desired in high-risk domains (such as health, autonomous vehicles), and recently required by law (GDPR regulations in EU). Several methods were developed to assure that feature, with LIME and SHAP among most robust approaches. In this paper we focus on the issue of interpretable decision making in time series data streams. In our previous work we developed causal rule-mining algorithm that provided contrastive explanations via rule-based notation. We work on extending this work to exploit strengths of process mining and Bayesian networks to better assure counterfactual explanations.