One of the still manual tasks involved in solving any problem by machine learning is feature engineering. This is often a time-consuming step and has a huge impact on the performance of the model. An important and useful property of machine learning models is the ability to estimate the uncertainty in their predictions. This is especially important in some specific areas, such as critical-safety applications or medicine. In this thesis, we first lay the theoretical foundations and later benchmark and compare a model that aims to tackle both challenges - the DeepGLM model. The model is a flexible extension of generalized linear models where the linear predictor is replaced by a deep neural network. By using a deep neural network and working in the Bayesian framework, the model is able to learn features and also estimate the uncertainty in the predictions. We benchmark the DeepGLM model on a real-world dataset from the insurance industry where we model the probability of a claim caused by tap-water accidents. We also test the capability of the model to learn features on a very small dataset. In our tests, the DeepGLM model compares favorably to other widely used methods in the insurance risk modelling. The DeepGLM model also achieved very good results in the feature-learning task, where it replaced the manual feature building. Furthermore, the model showed a good performance on a very small dataset consisting of only 16 samples, which is not commonly seen among neural network based models.     «

One of the still manual tasks involved in solving any problem by machine learning is feature engineering. This is often a time-consuming step and has a huge impact on the performance of the model. An important and useful property of machine learning models is the ability to estimate the uncertainty in their predictions. This is especially important in some specific areas, such as critical-safety applications or medicine. In this thesis, we first lay the theoretical foundations and later benchmark...     »