How Many Segments, Really? auditing a clustering claim

AI Tools & Strategy · Learning Design

The graded summative for Module 6 of my graduate Advanced Data Science course. Marketing wants to build campaigns for a fixed number of customer segments, and a junior analyst says a k-means elbow gives five. Your job is to audit how that number was produced and defend how many segments really exist.

What this is

A performance-based summative assessment for this module, worth 100 points. The interactive below takes about 30 to 45 minutes and covers the diagnosis and the defense, and it is self-scoring, so you can run it on your own. In the instructor-led version of the course the full assessment also includes the reproducible clustering analysis and the written segment profiles you submit, which is where most of the hours go, and each student is issued their own version of the dataset, seeded from their roster identifier, so the true number of segments differs from one submission to the next and the instructor can reproduce any of them.

Why you are doing this

Unsupervised results are hypotheses, not answers. With no labels to hold back, a cluster count cannot be read off a single elbow plot; it has to be earned through stability, internal validity, and interpretability. This assessment tests whether you can refuse to hand a stakeholder a clean number until the evidence actually supports it, the core discipline of honest unsupervised analysis.

What you are being assessed on

100 points total

15Cluster stabilityBootstrap co-clustering, not one elbow

15Agreement with ground truthAdjusted Rand (0 to 1 label-agreement score) vs held-out labels

14Rigor of the count argumentConverging evidence for the number

14Preprocessing and methodStandardize, PCA, more than one algorithm

14Actionable segment profilesDescribed for a marketer to use

14ReproducibilityRuns from a fixed seed

14Honesty about uncertaintyDo not overclaim a precise count

What you submit: in the instructor-led version this interactive covers the diagnosis and the defended count (57 of the 100 points), and the other 43 come from your reproducible clustering notebook, the agreement check against the held-out labels, and the written segment profiles you upload to the Module 6 assignment in the LMS. Working on your own, the interactive is self-scoring practice you can complete without an instructor.

SEGMENTATION REVIEW · MARKETING ANALYTICS

THE ASK

Your job: a colleague says the data has five customer segments. Work through the six steps to decide how many there really are, and defend it.

This simulation is an open educational resource by Michelle Blomberg, released under CC BY-NC-SA 4.0. Reuse, adapt, and share for noncommercial purposes with attribution, under the same license.