Final Project
The final project is a group project (groups of 2–3) worth 30% of the course grade: 20% for the in-class presentation and 10% for the final slide deck (the report). There will be five groups. Each group gives a 15-minute presentation (12 min talk + 3 min Q&A) on the last day of class, Friday 12/04. Every group member must present part of the talk.
You may choose either track:
- Applied analysis. Pick a real dataset and a question, and compare at least two methods from this course on it.
- Reproduction of a published analysis. Pick a published statistical learning paper that uses public data and repeat its analysis. Compare your results with the paper’s, and compare the paper’s method with at least one method from this course.
Key dates
| Date | What |
|---|---|
| Fri 10/30 | Groups formed. Email the instructor your group members. Anyone without a group by then will be assigned to one. |
| Fri 11/06, 11:59 PM | Proposal due (one PDF per group) |
| Fri 11/13, 11:59 PM | Last day to submit a proposal and still get feedback |
| Fri 12/04, 8:59 AM | Slides due (one PDF per group) |
| Fri 12/04, 9:00–10:50 AM | In-class presentations |
| Wed 12/09, 11:59 PM | Final report due: final slide deck (one PDF per group) |
0. Proposal (required, complete/incomplete)
One page (references may go on a second page). List the group members and a working title, then:
- Question. What you want to learn and why it matters. Reproduction track: the question the paper asks.
- Data. Name, source, and link. Number of observations, predictors, and outcome (if any). Confirm you already have access.
- Methods. The methods you plan to use and why they fit the question. Applied track: at least two course methods to compare, and how you will judge them (e.g. test-set AUC, cross-validated RMSE). Reproduction track: the paper’s method, the course method(s) you will compare it with, and the paper’s full citation.
Include references for every source you cite. Every reference must be real, and you must have read it.
The proposal is not part of the course grade, but it is required. It is marked complete if it has all three parts. An incomplete proposal gets comments and one week to resubmit. Proposals submitted after 11/13 get no feedback. A group without a complete proposal by the start of class on 12/04 loses 10% of its presentation grade.
1. Presentation (20%)
12 minutes of talk plus 3 minutes of questions. Follow this structure:
- Background — the scientific context, and why the problem matters.
- Question — the specific question you set out to answer.
- Method(s) — the data, any preprocessing, the methods, and how you evaluated them.
- Results — what you found. Use figures and tables, not raw output.
- Conclusions — what the results mean for the question.
Reproduction track: under Results, show what you reproduced and where it differs from the paper.
Each of the following is scored 0 (poor) to 5 (excellent):
| Item | What we look for |
|---|---|
| Background | Context is clear and motivates the question |
| Question | Specific and answerable with the data |
| Method(s) | Methods fit the question; evaluation is sound |
| Results | Clear figures/tables; results are correctly interpreted |
| Conclusions | Follow from the results; no overreach |
| Delivery | Clear slides, good pace, ends within 12 minutes |
| Q&A | Questions answered directly and correctly |
All members of a group receive the same grade.
2. Report (10%)
The report is your presentation slides, revised after the talk, with two extra slides at the end:
- Limitations — what weakens your conclusions (data, methods, sample size, assumptions).
- Next steps — what you would do with more time or more data.
Use the questions and comments from the 12/04 talks to fix or improve the other slides. End with a references slide. Submit one PDF per group. The deck is scored 0–5 on each of Background, Question, Method(s), Results, Conclusions, Limitations, and Next steps. Late reports follow the homework late policy.
Tips on choosing a project
- Applied track: pick a dataset where at least two methods from the course are plausibly relevant — that gives you something to compare. Public health data, genomic/single-cell, EHR-derived, or NHANES-style cohorts all work well. Places to look: UCI Machine Learning Repository, NHANES, Kaggle.
- Reproduction track: pick a paper whose data and analysis are described well enough to repeat. Check that you can download the data before you submit the proposal.
- Scope honestly: 12 minutes is about 8–10 slides. Plan the talk first, then do the analysis the talk needs.