Explanation using model-agnostic methods

Interpretable machine learning has become a hot topic as the complexity and capacity of machine learning models grow. However, these big and complex models, often referred to as black box models because their essence is difficult to understand, still boast high performance that cannot be given up. Model-agnostic methods are an approach to understand the predictive response of a black box model, instead of the response from the original dataset. This chapter first categorizes model-agnostic methods in two ways. They are first categorized by the goal of understanding such as analyzing marginal effect, measuring contribution of each feature, and finding a surrogate model. Then, they are categorized into global or local area of interest. This chapter discusses the focus and strategy in each of six domains. Representative outputs in six domains are demonstrated through a lab session using the statistical software R with the sophisticated package iml.