Concept and Principles of Causality
What is Causality?
Causality is a branch of knowledge that studies which events cause others and how much of an event (effect) is due to another (cause). It can be used to better understand how different variables interact within our system and how we can intervene to obtain more desirable outcomes.
Why Causality @ THN?
As a SaaS company that provides a personalization suite for hotel websites, The Hotels Network mission is to grow hotels’ direct channel. In order to do this, our first step is to understand user behavior inside the booking process and which factors drive their intentions. This way we can provide more insights, recommendations and a better experience to our customers.
Within the Data Science team at THN, we work on projects such as: user behavior prediction, time series based forecasts (prices, conversion rates, demand…) and websites’ content analysis through NLP. Recently, we dove into Causality as the means to understand how all the gears in our system click and interact with one another.
Our path to knowledge starts within the team. We prepare internal workshops on different topics, from introductory courses to more advanced state of the art technologies and algorithms.
This workshop is intended to give an overview on basic Causality concepts and principles that can be used to start a causality project. It gives an introduction to graphs, focusing on causality graphs and how to build them (Causal Discovery); do-calculus rules, used to simplify causal operations; and an overview of the Average Treatment Effect (ATE) and how to compute it. The ATE is the estimated effect that a treatment has, which allows us to estimate how the treatment improves the desired outcome by removing the bias from the specific samples that were selected to receive the treatment.
Most of the content has been taken from Brady Neal’s Causality video series, but hopefully I have managed to make it into a short introduction video. If after watching, you are more interested, make sure to check them out!
Extra Content
For those who felt like I rushed through the do-calculus slides, here are the backdoor and frontdoor adjustment formulas, explained using the do-calculus rules:
