Causal discovery and causal ML
Can the arrows be learned from data? The PC algorithm step by step, score-based and continuous search, and what no observational method can recover. Then machine learning in the service of causal questions, namely double/debiased ML, causal forests and uplift, invariance as a route to robustness, and what is known (and not) about causality in large language models.
55 min
After this lesson you can
- Run the PC algorithm by hand, and say which assumptions make its output correct
- State what observational data can and cannot identify about a causal graph, and name assumptions that go further
- Explain regularisation bias and how orthogonalisation and cross-fitting (double ML) remove it
- Describe causal forests, meta-learners and uplift modelling, and why evaluating them is hard
- Explain the idea of invariant risk minimisation and summarise the evidence against it
- Describe the current, disputed evidence on causal reasoning in large language models
In this lesson
- Learning a graph from independences: the PC algorithm
- Score-based and continuous search
- Beyond the equivalence class
- Machine learning for effect estimation: double ML
- Heterogeneous effects: whom to treat
- Invariance: causality as a route to robustness
- Causality and large language models
This lesson is part of Descent Pro
Descent Pro unlocks every lesson on the route, from base camp to the summit. Every plan starts with a free trial.