A Group Variable Importance Framework for Bayesian Neural Networks
Abstract: While the success of neural networks has been well-established across a variety of domains,
our ability to interpret these methods is still limited. Traditional variable importance
approaches in machine learning overcome this issue by providing local explanations about
particular predictive decisions - that is, they detail how important any given feature is to the
classification of a particular sample in the dataset. However, univariate mapping approaches
have been shown across many applications in the literature to generate false positives and
negatives in high-dimensional and collinear data settings. In this paper, we focus on the
slightly different task of global interpretability where our goal is to identify important groups
of variables by aggregating over collections of univariate signals to improve power and
mitigate false discovery. In the context of neural networks, a feature is rarely important on
its own, so our strategy is specifically designed to leverage partial covariance structures and
incorporate variable interactions into our proposed group feature ranking. Here, we extend
the recently proposed “RelATive cEntrality” (RATE) measure to the Bayesian deep learning
setting. We refer to this approach as the “GroupRATE” criterion. Given a trained network,
GroupRATE applies an information theoretic metric to the joint posterior distribution of
effect sizes to assess group-level significance of features. Importantly, unlike competing
approaches, our method does not require tuning parameters which can be costly and difficult
to select. We demonstrate the utility of our framework on both simulated and real data.
Submission Length: Long submission (more than 12 pages of main content)
Assigned Action Editor: ~Tongliang_Liu1
Submission Number: 857
Loading