Gang Fang's Blog

The new project I am working on at the defect coaching team requires a model-generated seller list, ranked by ROI. The model used for this job is a tree-based algorithm on https://github.com/uber/causalml. As I am studying this paper on Causal Inference https://www.jmlr.org/papers/volume11/spirtes10a/spirtes10a.pdf, the two models for Causal Inference are Causal Bayesian Networks and Structural Equation Models (SEMs). There aren’t tree-based algorithms or any other kinds of algorithms.

So a distinction needed to be made between two efforts: learning causal effects from data and causal machine learning. The first effort uses existing methods in the ML toolbox to identify causal effects and the question is what we can discover from the given data. The second effort attempts to create new approaches to ML based on principles of causal reasoning. This distinction is made based on the reading of this Microsoft Research’s overview.

To point out, the second effort above is causal machine learning but causal inference. Causal inference is a traditional field of statistics (even though its start dates back to probably 1990s and it isn’t old) while causal machine learning is more recent, a field that combines traditional machine learning and causal inference principles. Causal inference underlies multiple disciplines, such as econometrics, epidemiology and social science, and is most often approached through randomized experiments. But only causal machine learning seeks to address OOD generalization when causal inference doesn’t. Finally, while both rooted in the goal of understanding causality rather than mere correlations, causal machine learning can be seen as an extension of application of causal inference in the realm of machine learning.

New approaches to machine learning based on principles of causal reasoning provide a promising path forward. Guided by joint formal reasoning over observations and auxiliary information about data collection procedures or other domain knowledge, causal machine learning methods are grounded in the stable and independent mechanisms that govern the behavior of a system being modelled. As a result, these methods promise robustness to exogenous changes and accurate modelling of counterfactual or “what-if” scenarios that are core to scientific experimentation, understanding, and decision-making.