A0185
Title: Reluctant interaction inference
Authors: Guo Yu - University of California Santa Barbara (United States) [presenting]
Abstract: Additive models enjoy the flexibility of nonlinear models while still being readily understandable to humans. By contrast, other nonlinear models, such as neural networks, despite their popularity in the past decade, involve modeling interactions between features, which are much harder to interpret because they require describing how one or more features modulate the effect that other features are having on the response. Given the complexity of interpreting models with interactions, an analyst should be reluctant to move beyond an additive model unless it is truly warranted. The focus is on the high-dimensional setting and high standard of evidence for the existence of an interaction between two features. In particular, it is examined whether there is evidence that a linear interaction between features is at play after having fit a sparse additive model. This is formulated as a hypothesis testing problem in which an additive model acts as the null model. The p-values make use of recently developed machinery for selective inference after the group lasso.
