Program

Instructors:
Dr. Luis Enrique Sucar Succar.
Dr. Julio César Muñoz Benítez

The main objective of this tutorial is to introduce to the audience to the area of causal discovery, including the fundamentals and software tools. Initially, we will introduce graphical causal models, including the knowledge representation and inference techniques. Will present a comprehensive introduction to causal discovery methods equipping participants with the knowledge and skills to understand the fundamental concepts of causality and causal inference. Finally, we will have a practical session with some software tools for causal discovery in time series.

Part I: Fundamentals

In the first part of the tutorial, we will introduce causal graphical models and causal reasoning, with emphasis con causal discovery. Learning causal models from observational data is challenging, as in general we cannot recover a unique causal model, but a set of statistically equivalent models that are called a Markov equivalence class.

We will introduce some of the algorithms that can be applied to learning a causal model, and a technique that can try to obtain the most probable model within a Markov equivalence class.

We will present some applications, including discovering the effective connectivity in the brain, learning causal models from COVID data, and the impact of causal knowledge on predicting energy consumption price. This will provide an overview of the advantages of causal discovery in real-life scenarios.

Part II: Software Tools

The second part will focus on practical aspects, in particular in software tools for Causal Discovery in Time Series. Time series data have served as the basis for causal discovery in various fields of science. In this sense, data collected can provide very precise measurements at regular points of time. One of the main advantages of using time series is that the temporal order of the information can simplify the causal analysis.

Introduction to Tigramite

Tigramite is a python package for causal inference for time series. It allows to estimate causal graphs from time series datasets and to use graphs for robust forecasting, estimation, and prediction of direct, total, and mediated effects. Causal discovery is based on linear as well as non-parametric conditional independence tests applicable to discrete or continuous time series. We will use Tigramite to model a time series. This will help to understand and represent complex causal relationships and its flexibility in handling temporal dynamics. This will be helpful to describe the process of causal discovery on a synthetic time series.

We will explore the plotting functions of causal structure using Tigramite. This will allow to visualize the causal links between variables in time series data. These links are represented as directed edges in a graphical model, providing a clear visual indication of the cause-and-effect relationships. In this way, graphical models provide an intuitive and visual way to represent complex dependencies among variables in time series data.

Pre-requisites:

Basic knowledge of computer science is assumed, as well as programming experience (Python). Some prior knowledge on probability, machine learning and Bayesian networks will be useful, but not required.

Attendees should bring their own laptop for the practical part of the tutorial.

It is recommended to use Google Collab to follow the practical steps in time series analysis; the material is available at:

https://colab.research.google.com/drive/19notnYilbswEfWwjBJOzCruaduH5b3qq?usp=sharing.