CSCE 790 Section 1: Advanced Topics in Probabilistic Graphical Models

Meeting Time and Place: TTH 1140-1255, SWGN 2A15.

Bulletin Description: Topics in Information Technology. Reading and research on selected topics in information technology. Course content varies and will be announced in the schedule of courses by suffix and title. May be repeated for credit as topics vary.

Prerequisites: Graduate Standing in Computer Science, Mathematics, Statistics, or permission of the instructor.
Students from these three disciplines will be asked to contribute according to their background and interests. For example, a computer science student may be asked to explain the implementation of a Bayesian network learning algorithm in an R package; a statistics student may be asked to present a paper on the parametric variant of an inference algorithm, and a mathematics student may be asked to present a paper that describes the algebraic structure of a local computation framework.

Course Learning Outcomes. The overall goal of the course is to prepare students to carry out research in probabilistic graphical models. Specifically, by the end of this course, the student will be able to:

• Use the Hugin Bayesian network and influence diagram tool to construct Bayesian networks.
• Compare and contrast key algorithms for belief propagation in Bayesian networks based on local computation and relate them to a common algebraic foundation.
• Apply and explain variable elimination and several version of the junction tree methods at a detailed algorithmic level.
• Explain and justify the interventional model of causality.
• Apply Pearl's do-calculus of intervention, prove its properties, and describe its limitations in the parametric case.
• Compare and contrast the Markov properties of advanced probabilistic graphical models, such as chain graphs and ancestral graphs, and explain the uses of such models, with special focus on causal modeling.
• Explain the properties and limitations of structural learning algorithms for Bayesian networks, under the assumption of faithfulness (PC algorithm and its variants) and embedded faithfulness (FCI algorithm and its variants).
• Use a few key R packages for inference, learning, and causal modeling.

Reasonable accommodations are available for students with a documented disability. If you have a disability and may need accommodations to fully participate in this class, contact the Office of Student Disability Services: 777-6142, TDD 777-6744, email sasds@mailbox.sc.edu, or stop by LeConte College Room 112A. All accommodations must be approved through the Office of Student Disability Services.

Grading Policy

Time Allocation Framework

Lecture Log

Lecture Notes

1. Introductory PowerPoint Slides
2. Slides for the Introductory Examples

Presentations

Videos from the spring 2009 version of CSCE 582, which may be useful to catch up on background notions.

Lecture Notes from the spring 2009 version of CSCE 582, which may be useful for background notions. For example:

1. The pdf slides for Ch.1 [J07] have a good presentation of pointwise table operations in the context of probability computation
2. The pdf slides for Ch.2 [J07] define evidence as a vector of zeros and ones.
3. The transcript of notes of 2009-01-30 has a proof of the chain rules for Bayesian networks.

Notes on Non-Serial Dynamic Programming (NSDP) used on 2019-02-19.

This set of slides includes a presentation of variable elimination using relational algebra (slides 38-41), as used on 2019-02-19.