The Student Feedback Problem
How might we analyse student feedback while preserving privacy and identifying recurring issues?
The problem
Where students run into it today.
Thousands of responses. A summary nobody can act on.
Why it matters
Averages destroy information. A 3.4 out of 5 tells you nothing about the two specific things forty people wrote about.
Feedback is only honest when it is genuinely anonymous — which makes privacy a design constraint, not a compliance checkbox.
Recurring issues are hard to see one response at a time and obvious in aggregate. That gap is the entire opportunity.
Possible directions
Several credible approaches, no obvious winner. Choosing between them is part of the work.
Theme clustering over free text
Group comments by what they are about rather than by sentiment score, and surface clusters that repeat across terms.
Privacy by construction
Enforce minimum cluster sizes, strip identifying detail at ingestion, and make it structurally impossible to drill down to one person.
Change tracking
The valuable question is not what students said. It is whether what they said last term changed this term.
Technologies & difficulty
The approach is not known in advance. Expect a literature pass, prototypes that get thrown away, and results you have to argue for.
Difficulty describes the work, not the person. A level 04 problem is not closed to a first-year — it just needs a team and a longer runway.
Potential project
Campus Feedback Intelligence
A privacy-preserving analysis pipeline that clusters free-text feedback into recurring themes and tracks how those themes move over time.
Skills needed
Open roles & possible team
- Data
Pipeline, clustering, anonymisation guarantees
- Research
Methodology, validity, and what the clusters actually mean
- Frontend
Analysis interface and longitudinal views
Research questions
The things nobody in the group can currently answer.
What is the smallest group size at which a theme can be reported without risking re-identification?
Do sentiment scores add anything once you have themes?
How do you distinguish a real shift in opinion from a change in who responded?
Next steps
What the first two weeks would sensibly look like.
- 1
Write the privacy rules before any code, and treat them as acceptance tests.
- 2
Validate clusters by manually labelling a sample — if a human disagrees with the grouping, the grouping is wrong.
- 3
Build with synthetic or consented data only.