Phd thesis

Towards causal discovery for Earth system sciences

This work presents new causal discovery methods more suitable for the data regimes typically seen in Earth system sciences. I addressed some of the main challenges of non-additive, causal-insufficient, and spatio-temporal data as well as instantaneous causal relationships in dynamic deterministic systems. I also developed tools for data where interventions have occurred, but the variables and observations affected are not known. Furthermore, we present a methodology to estimate a latent causal representation for data from different interventional distributions to reduce the large discrete space on which causal discovery usually operates.

The thesis is structured in the following chapters:

  1. Introduction
  2. Non-additive data: The conditional mean embedding
  3. Non-additive data an no causal sufficiency: Latent noise imputation
  4. Robust convergent cross mapping
  5. Invariant causal prediction
  6. Discussion and conclusions

Here is a link to my phd thesis document:

And here are the slides of my phd Defense presentation

References