Course guide and study route

Course guide and study route#

Topics of Statistical Learning studies the statistical structure behind learning with incomplete information and inference after a data-dependent choice. The course combines mathematical derivations, small computational investigations, critical reading, and a focused research project.

Course information#

Item	Proposed arrangement
Instructor	Chi-Kuang Yeh · cyeh@gsu.edu
Course number	STAT 8XXX — number to be assigned
Level and credits	Advanced master’s / doctoral elective; 3 proposed credit hours
Teaching format	Lecture, seminar discussion, and guided computational investigations
Duration	15 teaching weeks; also suitable for independent study
Preparation	Graduate probability and inference, linear algebra, basic optimisation, statistical learning, and programming experience
Independent preparation	Approximately 6–9 hours per week during a taught offering
Offering information	Term, CRN, room, meeting times, and office hours to be announced

The public book is usable immediately. Course registration, dated deadlines, and institutional policies are supplied separately when an offering is confirmed.

Learning outcomes#

Specify the observed variables, latent variables, population target, and observation mechanism for a learning problem.
Derive PU and UU risk representations and explain the assumptions behind identification and stable estimation.
Distinguish noisy, complementary, partial, bag-level, and multiple-source labels, and recognise how their mechanisms differ.
Analyse identification, separation, comparison design, and dependence in preference models.
Formulate full-information and bandit problems, define an appropriate comparator, and derive a basic regret guarantee.
Distinguish ordinary, selective, and simultaneous inferential claims, including claims made after clustering.
Conduct and communicate a reproducible methodological investigation with a baseline, a sensitivity analysis, and uncertainty assessment.

A 15-week teaching route#

The chapters are learning units. The schedule allocates extra time to the assumptions and computations that benefit from discussion.

Wk	Question and material	Reading focus	Suggested milestone
1	Targets, risks, and observation mechanisms	The course’s model–target–mechanism framework	Computational readiness check
2	Forms of weak supervision	Natarajan et al.; optional Ishida and Ratner	Mechanism comparison
3	Positive–unlabelled learning	Elkan–Noto; PU risk derivation	Start Assignment 1
4	PU estimation and sensitivity	Kiryo et al.; selected Bekker sections	Prior and selection experiment
5	Two unlabelled mixtures	Lu et al.; inversion and conditioning	Assignment 1
6	Pairwise preferences	Turner–Firth; identification and estimation	Project proposal
7	Context and comparison design	Preference model extensions; optional DPO	Start Assignment 2
8	Online learning and regret	Selected Hazan sections	Work through the OGD bound
9	Bandit feedback	Selected Lattimore–Szepesvári chapters	Evaluate exploration
10	Duelling bandits	Yue et al.; winner criteria	Assignment 2
11	Selection and conditional inference	Lee et al.; truncated reference distributions	Reproduction checkpoint
12	Inference after hierarchical clustering	Gao–Bien–Witten	Start Assignment 3
13	Inference after k-means	Chen–Witten; optional data thinning	Audit algorithm choices
14	Research synthesis	Project-specific primary papers	Assignment 3; project clinic
15	Project presentations and discussion	Revisit the target and assumptions	Presentation; final report in the designated final period

Independent-study route#

Work through one chapter at a time. For each chapter, write down the model and target before reading the derivation. Run the supplied experiment unchanged, reproduce its main result, then alter one parameter or assumption. Answer at least one derivation question and one experimental question before revealing a solution. Keep a short record of what the experiment can and cannot establish.

For a narrower reading group, Chapters 1–5 form a weak-supervision block; Chapters 6–10 form a preference and sequential-learning block; Chapters 11–13 form an inference-after-selection block. Chapter 1 supplies notation for all routes, and Chapter 14 provides a shared project framework.

Proposed assessment for a taught offering#

Component	Weight	Timing
Three integrated assignments	45% total; 15% each	Weeks 5, 10, 14
Paper discussion and a two-page reading note	10%	Assigned during the term
Project proposal	5%	Week 6
Reproduction checkpoint	10%	Week 11
Final project report	20%	Designated final-assessment period
Project presentation	10%	Week 15

There is no proposed written midterm or final examination. A well-designed reproduction with one consequential extension is sufficient for the project. A new theorem or publication is not required. Details and rubrics are on the assignment and project pages.

Working principles#

Reproducibility. Record random seeds, software versions, preprocessing, data provenance, and an execution command. Separate training, method selection, and final evaluation. Report Monte Carlo uncertainty for simulated performance.

Collaboration. Discuss ideas and compare understanding. Identify collaborators and external sources, and write your own derivations, code, and interpretation. Any offering-specific collaboration policy takes precedence for assessed work.

AI tools. They may assist with coding, editing, or checking a line of reasoning when permitted by the offering. Record substantive use and verify the result. An apparently plausible derivation, citation, or executable program still needs independent checking. Do not put restricted or confidential data into an external service.

Accessibility and policies. Students should use the official accommodations and support channels at their institution. The final offering syllabus will include the applicable attendance, late-work, academic-integrity, accessibility, and support policies. This public study guide does not substitute for that document.