I’m a Ph.D. student at KIT Karlsruhe under the supervision of Dominik Janzing. You can reach me via e-mail under philipp.faller at partner.kit.edu.
My research focuses on Causality and how to learn causal models robustly from observational data.
Finding causal relationships in data is a fundamental task in science. Some scientific question only deal with observational aspects, such as “How high will the unemployment rate in Germany go next year?”. These questions are fundamentally different from questions incorporating changes, perturbations or interventions of the studied system. The latter incorporate the task to identify aspects of the data generating process that are unaltered by these interventions. A exemplary question from this regime would be “How high will the unemployment rate Germany go after introducing a minimum wage?” (where the minimum wage is the intervention in this example). We call these questions causal, as reasoning about interventions goes beyond classical observational statistics in that it requires to differentiate between cause and effect. Rigorously answering causal questions is subject of causal inference.
To reason about causality, one has to employ a model of the causal relationships between variables. Often in causal inference, this model is given a priori, e.g. through domain knowledge of an expert. Usually such prior knowledge is rather limited and its reliability is hard to asses.
Conducting (double blind) randomized control trial is considered the gold-standard of experimentally obtaining a causal model. But clearly conducting such a trial is very expensive. Moreover, sometimes it is not possible to conduct a randomized control trail either for practical or ethical reasons. Suppose, for example, we wanted to investigate the effect of smoking on lung cancer. We cannot randomly assign a lifelong smoking habit to the study participants. And if we expect to see an adverse effect of smoking on the participants life expectancy, it is hardly defendable to conduct such a study.
The focus of my Ph.D. project will not be to reason about given causal models, but to learn causal models from observational data. We will call this specific part of causal inference causal discovery. The ability to automatically (and robustly) discover causal relationships from data is crucial to make causality applicable to problems where no prior knowledge is available or experiments are not possible. Further, it could help uncover relationships that are too complex or involved for a human investigator to detect them. Especially in domains with high-dimensional measurements such as neural sciences or climate science.
Publictaions
A list of my publications can be found here.