B0753
Title: Best of both worlds: Combining interpretable transformation models with the flexibility of normalizing flows
Authors: Marcel Arpogaus - HTWG Konstanz (Germany) [presenting]
Abstract: In many real-world regression problems, data stems from complex distributions. As a result, a wide range of potential output values can arise for a given input vector, making it insufficient to predict the mean value solely. Density regression models allow a comprehensive understanding of the data by modelling the complete conditional probability distribution. Multivariate conditional transformation models (MCTMs) are combined, which have been recently introduced in the field of statistics, with state-of-the-art autoregressive normalizing flows (NF) from the machine learning community. The approach allows the leverage of the interpretability of MCTMs to model the marginal distributions of a multivariate response in the first step. In the subsequent step, the neural-network-based NF technique accounts for the complex and non-linear relationships in the data. To ensure computational efficiency, a masking technique derived from autoregressive flows is incorporated. The approach is compared with existing MCTMs and pure NF models on simulated and real data.
