B0586
Title: Distributional regression with neural networks in R
Authors: Lucas Kook - University of Copenhagen (Denmark) [presenting]
Lucas Kook - University of Copenhagen (Denmark)
Abstract: Prediction problems frequently feature complex response types and a mix of large tabular and non-tabular data. Classical regression models either break down under the computational load of processing such data or require additional manual feature extraction to render the problem tractable. Deeptrafo, an open-source implementation is presented for distributional regression with neural networks in the R language for statistical computing, which overcomes the above limitations by augmenting distributional regression models with deep neural networks. Models implemented in deeptrafo can handle univariate binary, ordinal, count, survival and continuous responses with autoregressive structures and uninformative censoring. The models are parameterized via neural networks and estimated via penalized maximum likelihood without assuming a parametric family for the conditional outcome distribution. Special cases include neural network-based versions of linear, logistic, Weibull and Cox regression. Using deeptrafo, the data analyst can trade off interpretability and flexibility by supplying custom neural network architectures and smoothers for each term in an intuitive formula interface. It demonstrates how to set up, fit and work with deeptrafo on a real-data application, including in-built ensembling, cross-validation and visualization methods.
