Discriminative and Generative Models in Causal and Anticausal Settings

摘要

Having knowledge about the real underlying causal structure of a data generation process has various implications for different machine learning problems. We address the idea of causal and anticausal learning with respect to a comparison of discriminative and generative models. In particular, we conjecture the hypothesis that generative models perform better in anticausal problems than in causal problems. We empirical evaluate our hypothesis with different real-world data sets.