Partially Interpretable Neural Networks for High-Dimensional Extreme Quantile Regression: With Application to U.S. Wildfires

• Activity Number: 112 - Risk Analysis in Environment and Health
• Type: Contributed
• Date/Time: Monday, August 8, 2022 : 8:30 AM to 10:20 AM
• Sponsor: Section on Risk Analysis
• Author(s): Jordan Richards* and Raphael Huser
• Companies: King Abdullah University of Science and Technology and King Abdullah University of Science and Technology (KAUST)
• Keywords: spatio-temporal extremes; extreme value analysis; environmental risk; point process; quantile regression; neural networks

Abstract

Quantile regression is a powerful tool for modelling risk. If our interest lies in quantifying risk associated with extreme weather events, we may want to estimate conditional quantiles outside the range of data by using parametric models with parameters represented as functions of covariates. Classical approaches for extreme quantile regression use linear or additive functions, and such approaches suffer in either their predictive capabilities or computational efficiency. Neural networks can capture complex relationships between variables and scale well to high-dimensions. However, statisticians may choose to forego their use as a result of their “black box" nature; although they give accurate prediction, it is difficult to use them for statistical inference. Inspired by recent focus in literature on “explainable AI", we propose a framework for performing extreme quantile regression using partially interpretable neural networks. Model parameters are treated as combinations of neural networks and readily interpretable linear and additive functions. We predict extreme quantiles of burnt area and occurrence probabilities for wildfires over the entire contiguous U.S.