Machine Learning for Civil and Environmental Engineers

8 Causal Discovery and Causal Inference

There is no result in nature without a cause. Understand the cause, and you will have no need for experiments.

Leonardo da Vinci

Synopsis

The vast majority of this textbook is devoted to exploring, creating, applying, and assessing machine learning models. While this chapter continues this trend, yet unlike the rest, this chapter dives into the data in the hope of discovering the hidden mechanism(s) that generate the observations we are interested in.¹ My goal in this chapter is to divert your attention, for a bit, from the linear process of collecting data → run algorithms → arrive at a blackbox/whitebox ML model and toward collecting data → identify the data generating process² (DGP) → arrive at a causal ML → infer the effects of manipulation(s)/treatment(s). By now, you might be wondering. What is a DGP? What is a causal model? What are manipulations and treatments?

For this, allow me to first introduce some of the big ideas behind causality, its principles, and how we can integrate those into a ML paradigm.

8.1 Big Ideas Behind This Chapter

A fundamental pursuit for us is to understand how a phenomenon comes to be. As engineers, we design and conduct experiments to uncover cause(s) and effect(s)³ behind a phenomenon we happen to be interested in. Our thinking process is simple, manipulate the system in one way to observe how the system responds. The outcome of such manipulation can then be thought of as the result that reflects ...