A0912
Title: Approximation principle for domain generalization
Authors: Gloria Buritica - AgroParisTech (France) [presenting]
Sebastian Engelke - University of Geneva (Switzerland)
Abstract: In climate science studies, predicting the future impact of climate variables is crucial. In a climate change context, environmental variables like temperature and precipitation amounts increasingly reach extreme levels, highlighting the need for accurate predictions in these extreme scenarios. Machine learning algorithms are popular for regression due to their strong predictive power on test points, which are well-represented during training. Their effectiveness comes from their ability to interpolate data without strict assumptions. However, these methods often fail to generalize well to underrepresented test points, struggling to predict extreme or rare events, as machine learning algorithms are not tailored to extrapolate. A median regression method is presented that improves predictions when covariates take extreme values. The approach is based on a regression extrapolation principle that models the response variable as covariates reach their highest levels. It works under minimal non-parametric restrictions, allowing satisfactory generalization results to be achieved.
