Out-of-distribution Generalisation via Diluted Causality

Homer | Apr 1, 2025 min read

đź“„ Paper accepted at AISTATS 2025

Description

The term diluted causality comes from Bühlmann (2018), inspired by a suggestion from Edward George. It reflects the notion that while causal mechanisms are expected to be invariant under all possible interventions, a weaker form of invariance—diluted causality—can be more desirable in practice. This weaker notion restricts the set of interventions to those most relevant for predictive generalisation.

This project investigates the interplay between causality, invariance, and out-of-distribution (OOD) generalisation. It builds upon influential work by Peter Bühlmann, Nicolai Meinshausen, Jonas Peters, Dominik Rothenhäusler, Rune Christiansen, and others.

We especially recommend reviewing these foundational papers to understand the core concepts connecting causality and robustness:

This project offers a concise introduction to this learning framework and highlights our contribution to the field. For a quick primer, check out the following:

👉 Introduction to out-of-distribution generalization from a causal perspective