A QUEST for Model Assessment: Identifying Difficult Subgroups via Epistemic Uncertainty Quantification.

Abstract

Uncertainty quantification in machine learning can provide powerful insight into a model's capabilities and enhance human trust in opaque models. Well-calibrated uncertainty quantification reveals a connection between high uncertainty and an increased likelihood of an incorrect classification. We hypothesize that if we are able to explain the model's uncertainty by generating rules that define subgroups of data with high and low levels of classification uncertainty, then those same rules will identify subgroups of data on which the model performs well and subgroups on which the model does not perform well. If true, then the utility of uncertainty quantification is not limited to understanding the certainty of individual predictions; it can also be used to provide a more global understanding of the model's understanding of patient subpopulations. We evaluate our proposed technique and hypotheses on deep neural networks and tree-based gradient boosting ensemble across benchmark and real-world medical datasets.