Lab 23 — Project Kickoff: Proposal Workbook

Session: Week 12 · Session 23 · 90 min CLO: CLO-2, CLO-3 Difficulty: Intermediate

Learning objectives

By the end of this lab you can:

1. Turn a vague interest into a specific, answerable data question.
2. Choose a dataset that actually exists and is licensed for use.
3. Define a success metric before building anything.
4. Draft the first version of the project proposal (due Session 26).

Problem statement

Your final project (teams of 2–3, see projects/README.md and assessment-plan.md) starts now. This lab is a workbook: by the end of this session each team has a one-page proposal draft with a dataset verified to load, a target variable, and a success metric. The real proposal (due W13 S26) will be this draft, polished.

Dataset requirements

Any public dataset your team can actually load today. Options that load with one line: seaborn built-ins (penguins, tips, flights, diamonds, titanic), or anything reachable via a keyless API (Open-Meteo pattern from Lab 11). Verify it loads in this session — an unloadable dataset is a failing proposal.

Step-by-step tasks

1. Team + question (15 min): write one sentence each:
2. the question (must be answerable with data),
3. who cares about the answer,
4. what decision the answer would inform.
5. Dataset check (20 min): load the dataset; print shape, info(), head(). Confirm: it exists, loads offline (or caches), and you know its license/provenance (write it down).
6. Framing (15 min): write down:
7. target variable (what you predict/describe),
8. feature candidates,
9. supervised or unsupervised? regression or classification?
10. Success metric (10 min): one line: "success = test MAE ≤ X" or "success = CV accuracy ≥ Y" — pick a number you can check in Lab 24.
11. Risk list (10 min): list the top 3 things that could go wrong (bad data, metric too optimistic, time) and one mitigation each.
12. Repo setup (20 min): in your course repo, create projects/final/ with: README.md (the draft proposal), data/ (or a data_loader.py that fetches the dataset), and a first eda.ipynb with the task-2 verification. Commit and push.

Starter code

# Project Proposal (draft)

**Team:** <names>   **Date:** <date>

## Question
<one sentence: answerable with data>

## Audience & decision
<who cares, what decision>

## Dataset
- Source: <name / URL>
- License / provenance: <...>
- Shape: <rows, cols>   Verified loads: yes/no

## Framing
- Target: <...>
- Features: <...>
- Type: supervised/unsupervised · regression/classification

## Success metric
- <one number, checkable>

## Top risks
1. <risk> -> <mitigation>
2. ...
3. ...

Expected output

• projects/final/README.md committed with all 7 sections filled.
• projects/final/eda.ipynb committed showing the dataset loads (shape/info/head cells executed).
• A one-sentence status: "proposal draft done, pending instructor review".

Questions

1. Why must the question be answerable with the data you have? Give an example of a great question and an unanswerable one.
2. What is the difference between the target and a feature?
3. Why define the success metric before building the model?
4. Why does dataset provenance/license matter for a graded project?
5. Your metric is "CV accuracy ≥ 90%". What is one way that number could be misleading?

Challenge task

Write the threats-to-validity paragraph (5–6 sentences): under what conditions would your conclusion be wrong? Include at least one data problem (missingness, bias in sampling) and one modeling problem (overfitting, leakage). This paragraph becomes part of the final report (Session 32).

Marking rubric

Criteria	Max	Notes
Question + audience + decision	4	specific, answerable
Dataset verified loading	4	shape/info/head executed
Provenance/license recorded	2	stated in README
Framing (target/features/type)	3	correct terminology
Success metric with number	3	checkable
Risks with mitigations	3	3 risks, 3 mitigations
Repo structure + commit	4	README + eda.ipynb pushed
Answers to questions	3	Q1, Q3, Q5 correct
Challenge: validity paragraph	4	≥2 problem types
Total	30