Supervised machine learning was never aimed for decision-making, yet machine learning models are increasingly used for decision-making in critical domains like healthcare and governance. The supervision loss optimizes
for prediction assuming that all relationships stay the same, whereas decision-making involves intervening to change exactly those relationships. Correspondingly, two directions are gaining popularity: explanation methods that allow people to check whether a model learns the right features, and training methods that correct correlational biases learnt by the model.

In this talk, I will show how both model explanation and real-world decision-making are fundamentally causal problems, and propose a unified framework to address them. Specifically, these problems are special cases of forward (what will happen if I do this?) and reverse causal questions (why did something happen?). In model explanation, we are interested in the backward question for the model's output; the forward question is trivial. In decision-making, the focus is on the forward question; the backward question is often intractable. Using the proposed causal methods, I will describe case studies on medical risk prediction, face classification, and user ranking for recommender systems. This work has led to two open-source libraries, DiCE and DoWhy, that are being widely used for applying causal reasoning to new applications.