Proxy Model Explanations for Time Series RNNs

While machine learning models can produce accurate predictions of complex real-world phenomena, domain experts may be unwilling to trust such a prediction without an explanation of the model’s behavior. This concern has motivated widespread research and produced many methods for interpreting black-box models. Many such methods explain predictions one-by-one, which can be slow and inconsistent across a large dataset, and ill-suited for time series applications. We introduce a proxy model approach that is fast to train, faithful to the original model, and globally consistent in its explanations. We compare our approach to several previous methods and find both that methods disagree with one another and that our approach improves over existing methods in an application to political event forecasting.