Abstract
Traditional machine learning techniques focus on developing predictive models that have the sole purpose of obtaining a high degree of accuracy on a given data set. These types of models exploit any type of association between the input and target variables that may increase the performance. However, in practice, the training and test distribution often differ significantly, resulting in unreliable and failing models. The key to learning generalizable models that work in a broad range of environments (and that are not affected by small changes in the test distribution) lies in learning causal predictive features. However, learning causal models under changing environments and in systems with hidden confounders is an unsolved problem and is directly connected to the generalisation gap. In this project, we aim to use the novel framework of causal machine learning to develop algorithms that can handle changing environments. More specifically, this project focuses on learning invariant and causal representations from data using causal machine learning. The results are models that are proven to be more generalizable, can cope with interventions, and are able to extract interpretable causal relations directly from data.
Researcher(s)
Research team(s)
Project type(s)