What this assignment does
The report compares logistic regression, support vector machines and random forests for early breast cancer diagnosis, evaluated by cross-validation with threshold analysis.
Its conclusion is the interesting part: the SVM recorded the best single scores and the report recommends logistic regression anyway, on grounds of interpretability, calibration and the stakes of the decision.
How it is built
Literature, dataset and methodology, results, critical discussion, recommendations. The threshold analysis in the results is what makes the later argument about calibration possible — the structure is load-bearing.
Models are compared on identical folds and identical metrics, so differences are attributable to the model rather than the evaluation.
Annotated extracts
Three passages, each with a note on why it earns marks and what a weaker answer does in the same place — enough to teach the move, not enough to stand in for the assignment.
Although the SVM recorded the best single scores, logistic regression is arguably the most appropriate model for clinical decision support. Its performance is statistically indistinguishable from the SVM, its coefficients can be inspected directly, and its probabilities are well calibrated, which made the threshold analysis above possible. This is consistent with Rudin's (2019) position that complex models are hard to justify in high-stakes settings when a transparent one performs equally well.
Why this marks well
The report declines to recommend the winner, and every clause of the justification does work: statistically indistinguishable (so nothing is lost), coefficients inspectable (so a clinician can audit it), probabilities calibrated (so the threshold can be tuned to the clinical cost of a miss).
“Statistically indistinguishable” is the load-bearing phrase. Without it the argument would be trading accuracy for interpretability; with it, there is no trade at all, and the recommendation becomes obvious rather than principled-but-costly.
Rudin (2019) is exactly the right citation — her argument is specifically that high-stakes decisions should not use black-box models when interpretable ones match them. The source is load-bearing rather than decorative, and it is cited after the student's own reasoning rather than instead of it.
What a weaker answer does here
A weaker answer reports that the SVM achieved the highest accuracy and recommends it. The analysis is competent and the recommendation ignores everything about the setting in which the model would be used.
The SVM achieved the best balance, with the highest cross-validated score. Lowering the classification threshold further increases sensitivity at the cost of specificity.
Why this marks well
Reporting the threshold trade-off rather than a single accuracy figure shows the student understands that a classifier is a family of operating points, not one number.
In a diagnostic setting this is the whole question. A false negative means a missed cancer and a false positive means an unnecessary biopsy — these are not symmetric harms, so the threshold is a clinical decision rather than a statistical default at 0.5.
It also sets up the interpretability argument: only a well-calibrated model lets you move the threshold meaningfully, which is why calibration appears in the recommendation.
What a weaker answer does here
A weaker answer reports accuracy at the default threshold for all three models and compares those numbers. The comparison is valid and answers a question nobody in a clinic would ask.
Distance-based and margin-based methods such as the SVM require feature scaling, which was applied within cross-validation folds rather than before splitting.
Why this marks well
Scaling inside the folds rather than before the split is a one-line detail that separates students who understand data leakage from students who have followed a tutorial.
Scaling on the full dataset first lets information from the validation fold influence the training fold, which inflates every score that follows. The report would still have produced plausible numbers — just quietly wrong ones.
Stating it explicitly also tells the marker the comparison is trustworthy, which is what licenses the whole results section.
What a weaker answer does here
A weaker answer normalises the dataset, then splits it, then reports strong results. Nothing looks wrong, the leak is invisible in the output, and every figure in the report is slightly optimistic.
What makes this a strong answer
It recommends against the best-scoring model and makes the case watertight by first establishing that no accuracy is being given up.
It treats classification as a choice of operating point in a setting where the two error types carry very different costs.
The methodology closes the leakage route that would have invalidated the comparison, and says so.
Request the full document
Request the full document
The extracts above are the teaching material and they stay on this page. The complete 2,050-word document is sent person to person, by a human, after a short conversation about what you are working on — so that we can point you at the right parts of it rather than hand over a file and wish you luck.
Both links below already name this sample (DSC-02), so you do not have to describe it.
WhatsApp · +44 7537 126 976 · Email · mia@globalprojectshelp.com
Mon–Sat, 9–22 GMT. No form, no mailing list, no download gate.
Full documents are never published, hosted or linked on this site.
Using this sample honestly
This is a worked example. It sits in the same place as a worked example in a textbook or a past paper with an examiner’s report attached: read it, take the move, then write your own paragraph without it open in another window.
Submitting any part of this document as your own work would be plagiarism — the same offence as copying from a journal article, and detected the same way. If you draw on it, cite it. If a deadline has got to the point where submitting someone else’s paragraph feels like an option, say so when you message us; that is a problem with a legitimate answer and we would rather work on it than watch it happen.
Our full position, including how to declare mentoring support to your institution, is on the academic integrity policy page.
Where to go next
- More MSc Data Science samples The full set of annotated model answers for this course.
- All samples Every annotated model answer in the library, with discipline filters.
- How requesting works What you receive, what you do not, and the conditions attached.
- Academic integrity policy What we do, what we will not do, and how to declare mentoring support.
Model answer written by a Global Projects Help mentor · © Global Projects Help
Published 8 October 2026 · Questions? mia@globalprojectshelp.com