Interpretable Machine Learning -- A Brief History, State-of-the-Art and Challenges

摘要: We present a brief history of the field interpretable machine learning (IML), give an overview state-of-the-art interpretation methods and discuss challenges. Research in IML has boomed recent years. As young as is, it over 200 years old roots regression modeling rule-based learning, starting 1960s. Recently, many new have been proposed, them model-agnostic, but also techniques specific to deep tree-based ensembles. either directly analyze model components, study sensitivity input perturbations, or local global surrogate approximations ML model. The approaches state readiness stability, with not only proposed research, implemented open-source software. But important challenges remain for IML, such dealing dependent features, causal interpretation, uncertainty estimation, which need be resolved its successful application scientific problems. A further challenge is missing rigorous definition interpretability, accepted by community. To address advance field, we urge recall our interpretable, data-driven statistics (rule-based) ML, consider other areas analysis, inference, social sciences.