ECE 69500 - Probabilistic Causal Inference

Course Details

Lecture Hours: 3 Credits: 3

Areas of Specialization:

  • Communications, Networking, Signal & Image Processing
  • Computer Engineering

Counts as:

Normally Offered:

Fall - odd years

Campus/Online:

On-campus only

Requisites:

ECE 302

Requisites by Topic:

Probabilistic methods in electrical and computer engineering

Catalog Description:

Causality is a fundamental concept, which plays critical role in several areas today including machine learning and data science. In this course, we are going to learn the tools for modeling probabilistic causality. We will learn causal graphs and how they can be used for estimating interventional queries. We will also cover learning causal graphs from observational as well as interventional data. Finally, we are going to briefly discuss the notion of counterfactuals. The course also has a research component and students are expected to develop or outline a novel algorithmic idea/solution or a new application of the tools they learn in the class to their research, which they will describe in their project presentation and project report.

Required Text(s):

None.

Recommended Text(s):

  1. Causality: Models, Reasoning and Inference , 2nd Edition , Pearl, J. , Cambridge University Press , 2009 , ISBN No. 978-0521895606
  2. Causation, Prediction, and Search , 2nd Edition , Spirtes, P., Glymour, C., Scheines, R. , MIT Press , 2001 , ISBN No. 9780262527927
  3. Probabilistic Graphical Models, Principles and Techniques , Koller, D., Friedman, N. , MIT Press , 2009 , ISBN No. 9780262013192
  4. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference , 2nd Edition , Pearl, J. , Elsevier , 1988 , ISBN No. 9780934613736

Lecture Outline:

1 Topic
1 A brief history of causality, independence, conditional independence, independence models
2 Graphoid and semi-graphoid axioms, undirected and directed graphs as dependency models
3 Bayesian networks, factorization d-separation, global Markov property, local Markov property
4 Causal graphs, doing vs. seeing, interventions and causal effect, do calculus
5 Presence of latent variables, c-components
6 ID algorithm, soundness and completeness
7 Learning causal graphs (no latents), observational equivalence class, essential graph
8 Constraint-based learning, faithfulness assumption PC algorithm
9 Score-based learning; GES algorithm
10 Learning causal graphs with latent variables, ancestral graphs, maximal ancestral graphs
11 Observational equivalence class with latent variables inducing paths, discriminating paths, partial ancestral graphs, FCI algorithm
12 Bivariate causal discovery; linear models, additive noise models
13 Independence of cause and mechanism, entropic causality
14 Experimental design for causal discovery, challenges, graph theory basics
15 Connections to graph coloring, separating systems and graph separating systems
16 Interventional equivalence classes, learning I-MECs (no latents) and I-PAGs (with latents)
17 Introduction to counterfactuals

Assessment Method:

Research report and presentation, exam