A0320
Title: Neural distributional regression models
Authors: Benjamin Saefken - Clausthal University of Technology (Germany) [presenting]
Abstract: Users of neural regression models have a tendency to neglect distributional aspects of the data. This is often because appropriate frameworks are not available. Neural Additive Models for Location, Scale and Shape (NAMLSS) provide a framework that allows the modelling of distributional aspects of the target data. This approach offers many advantages for the user, such as accurate prediction intervals. In comparison to classic statistical regression models, their neural counterparts allow for the possibility of incorporating non-tabular data. However, for a statistician, this is only of interest if it is done in an interpretable fashion. Otherwise, inferential techniques are not sensible. An approach for specifically incorporating images in a comprehensible way is proposed based on embedding spaces. Results on particular applications are promising. The downside is that the method is not easily generalizable to other settings and data sets and not as mathematically rigid as common regression models for tabular data.
