Choosing a question, finding data, and planning honest evaluation
Introduction to Data Science and Machine Learning · Md Mahbub E Noor · University of Barishal
question -> data -> baseline -> models -> honest test -> report
One sentence: what you predict or discover, and for whom it would be useful.
Enough rows (a few hundred at least), legally and ethically obtained, usable offline.
A metric that matches the cost of mistakes, and a baseline to beat.
| Idea | Data | Type |
|---|---|---|
| Bangla or Banglish SMS spam | Messages you collect with permission, anonymised | Classification |
| Wine, cancer or digits deep-dive | Built into scikit-learn | Classification |
| Diabetes progression | Built into scikit-learn | Regression |
| Student study habits | A short anonymous survey of classmates | Regression or clustering |
| Local prices or weather | Records you obtain as a CSV file [check the source] | Regression |
| Your own idea | Discuss with the instructor first | Any |
Each milestone is a short check-in. Problems found early are easy to fix.
| Section | What to write |
|---|---|
| Title and team | Names and student IDs |
| Question | One sentence: what you predict or discover |
| Data | Source, size, features, label; how you get it; privacy plan |
| Type | Classification, regression or clustering |
| Evaluation | Main metric and why; baseline; split and cross-validation plan |
| Risks | What might go wrong, and your backup plan |
| Part | Weight | What earns marks |
|---|---|---|
| Proposal | [__%] | Clear question, realistic data plan |
| Method | [__%] | Correct split, baseline, pipeline, cross-validation |
| Results | [__%] | Honest test score, right metric, error analysis |
| Report | [__%] | Clear, one structure, limits stated |
| Presentation | [__%] | Explains choices and results in 5 minutes |
One clear question beats five vague ones. Shrink the scope early.
A score means nothing without a simple guess to compare with.
Choose everything with cross-validation; test once, at the end.
Permission, anonymisation, safe storage. No personal data in submitted files.
[Course AI policy]. You must be able to explain every line of your code.
[Team size and rules]. Every member presents part of the work.
Proposal due: [date]
Pick a question and check that the data really exists and is usable offline.
Load the data and make two charts. Note the size, the label and the class balance.
Write the one-page proposal using the template. Submit by [date].