| As machine learning models increase in complexity, the human ability to understand and interpret decisions made by those models has not been able to keep up. The usefulness of white-box analysis techniques, exposing the internal state of models, is limited to relatively simple models and has trouble with complex models, such as deep neural networks. Recently, black-box machine learning analysis techniques offer model independent insights into the decision making process of machine learning models. In order to quickly and effectively gain insights from those techniques, visual analytics emerges as a powerful set of tools. We use visual analytics to explore both global and local means of explaining and understanding predictive models via black-box techniques. We then propose the Model Diagnostic workflow that uses aggregated instance-level explanations to overcome problems of fully global or local methods. That is, by avoiding global aggregates, finer details of the decision making process are retained, while going beyond individual instances, analysts are not overwhelmed by the quantity of instances to inspect. Finally, we show that the Model Diagnostic workflow can not only help improving the models themselves but offers insights about flaws in the input data, thus helping with the task of feature engineering. |
