Week 8 · Hands-on foundations
How can a group turn an interesting topic into a feasible and evidence-based data science project?
The current Advanced Data Science materials devote Week 8 to PBL instructions and a worked diabetes project, and schedule the first presentation in Week 10. Intermediate Data Science uses the same two-week preparation interval. This tutorial keeps that familiar structure while reducing the worked example to a reusable path: question, data, exploration, baseline, neural comparison, held-out evidence, and limitations.
This chapter condenses materials taught in the 2023–2025 Advanced Data Science course into a smaller core path. Optional depth remains available in the companion notebook.
Use the explorer and three short Python examples before changing a longer model.
Add one ablation or three-seed comparison that tests whether the apparent improvement is stable and attributable to the proposed model change.
Core ideas
A project begins with a population, outcome, and intended use. A model name is not a research question. The group should be able to state what will be predicted, for whom, with which available information, and why the result matters.
A credible project connects provenance, exploratory analysis, a simple baseline, a valid split, a comparable neural model, held-out results, representative failures, and limitations. Removing one link weakens the claim.
Group members may specialize, but every member must understand the complete question-to-evidence path. The presentation should identify contributions without exposing private personal details on the public course page.
Mathematical intuition
This is a planning relationship, not a numerical scoring formula. It reminds a group that a more complex model cannot compensate for an unclear question, invalid split, or missing limitation.
Identify the input, learned quantity, output, and aggregation. Re-create the relationship with a tiny hand-checkable example before using a library layer.
Interactive explorer
Predict the direction of change before moving a control.
Change a control to inspect the relationship.
Hands-on path
Minimal Python
Each example is independent. Explain its output before copying it into a larger workflow.
rows = 120
train = int(rows * 0.70)
validation = int(rows * 0.15)
test = rows - train - validation
print(train, validation, test)Predict the result, then click Run.
baseline_mse = 1.00
neural_mse = 0.72
relative_reduction = (baseline_mse - neural_mse) / baseline_mse
print(round(relative_reduction, 2))Predict the result, then click Run.
parts = {'question':1, 'data':1, 'baseline':1, 'held_out':1, 'failures':1, 'limitations':0}
missing = sorted(name for name, ready in parts.items() if not ready)
print(sum(parts.values()), len(parts), missing)Predict the result, then click Run.
Failure checks
What goes wrong: The group chooses an architecture before defining a feasible question.
Check: Ask whether the target, observation unit, intended use, and baseline fit in one sentence.
What goes wrong: The final test set is repeatedly used to choose preprocessing or architecture.
Check: Keep the test set closed until the model and decision rule are frozen.
What goes wrong: The notebook contains figures and metrics but the presentation never states what they support.
Check: Require every main slide to connect one claim to one visible piece of evidence and one limitation.
Check your understanding
Answer in your own words before opening each panel.
Standard answer: A strong answer connects the central idea to a visible computation and a held-out or shape-based check. This tutorial keeps that familiar structure while reducing the worked example to a reusable path: question, data, exploration, baseline, neural comparison, held-out evidence, and limitations.
Standard answer: A project begins with a population, outcome, and intended use. A model name is not a research question. The group should be able to state what will be predicted, for whom, with which available information, and why the result matters.
Standard answer: A credible project connects provenance, exploratory analysis, a simple baseline, a valid split, a comparable neural model, held-out results, representative failures, and limitations. Removing one link weakens the claim.
Standard answer: Group members may specialize, but every member must understand the complete question-to-evidence path. The presentation should identify contributions without exposing private personal details on the public course page.
Standard answer: This is a planning relationship, not a numerical scoring formula. It reminds a group that a more complex model cannot compensate for an unclear question, invalid split, or missing limitation. It does not by itself prove useful behavior on unseen intended-use cases.
Standard answer: The visible calculation and interpretation should change together. Predict the direction first, then use the result to correct your mental model.
Standard answer: Write one sentence naming the prediction target, observation unit, intended use, and non-use.
Standard answer: Ask whether the target, observation unit, intended use, and baseline fit in one sentence.
Standard answer: Record data and split assumptions, input/output shapes, settings, the baseline, held-out metrics, representative failures, and the decision supported by that evidence.
Standard answer: Week 9 adds convolution and a compact CNN workflow. Week 10 is reserved entirely for the first PBL presentations, two weeks after this tutorial.
Terminology
Go further
Next: Week 9 adds convolution and a compact CNN workflow. Week 10 is reserved entirely for the first PBL presentations, two weeks after this tutorial.