Incorporating Expert Interpretation and Reasoning to Guide Model Selection in Machine Learning

Virtual Dissertation Defense

ABSTRACT:

In using machine learning to train predictive models, training data often under-specify solutions due to limited sample size. When selecting a model for deployment, we must consider metrics beyond training loss to identify potential pitfalls of the learned model and seek opportunities to correct them. To address model underspecification, in this thesis, we develop several methods that leverage domain knowledge during model selection.

First, to select among solutions, one must understand the learned model. Our approach, Shapley Flow, takes a user defined causal graph on the features as input and summarizes the attribution to model prediction along the causal edges. Shapley Flow unifies three widely used Shapley-value based model interpretation methods and elucidates the need to consider the data generation procedure, capturing both the direct and indirect impact of features. Second, domain knowledge is not only useful in model interpretation, it is also crucial in training. If one knows what features experts rely on, one can incorporate such knowledge to avoid using proxy features that are spuriously correlated with the outcome. We propose a novel regularization technique, Expert Yield Estimate (EYE), to learn a credible model, one that is both accurate and is aligned with domain experts. Finally, we connect credible learning with shortcut learning, identifying sufficient assumptions for credible models to eliminate the dependence on spurious correlation. In this process, we extend the EYE penalty to work with nonlinear models and work on tasks with domain knowledge not expressed on the input space. By leveraging domain knowledge, our proposed approaches help build trustworthy systems that can be safely applied in practice.