NODE-GAMLSS: Interpretable Uncertainty Modelling via Deep Distributional Regression

Ananyapam De, Anton Frederik Thielmann, Benjamin Säfken

Keywords: distributional regression, uncertainty modelling, neural regression, additive models

TL;DR: Interpretable uncertainty modelling using distributional regression with decision ensembles

Abstract: We propose NODE-GAMLSS, a framework for scalable uncertainty modelling through deep distributional regression. NODE-GAMLSS is an interpretable attention based deep learning architecture which models the location, scale, and shape (LSS) dependent on the data instead of only the conditional mean enabling us to predict quantiles and interpret the feature effects. We perform a benchmark comparison based on simulated and real datasets with state-of-the-art interpretable distributional regression models, demonstrating the superior quantile estimation, accuracy and interpretability.

Submission Number: 14