Recap and final exam
Ten modules, three families of methods, one shared discipline of reading. Here is the course condensed — module by module first, then through the threads that run across it.
The course at a glance
| Module | Core takeaway |
|---|---|
| 1. Searching without labels | The goal moves from predicting to revealing structure; no unique answer; scaling decides everything |
| 2. k-means | Alternating assignment and update; minimizes inertia; assumes spherical clusters |
| 3. Choosing k | Inertia always falls: read its slowdown (elbow); silhouette compares across k |
| 4. Hierarchical | The whole tree first, the cut afterwards; Ward by default; quadratic cost |
| 5. DBSCAN | A cluster is a dense region: free shapes and explicit noise; eps via the k-distance graph |
| 6. PCA | Axes of maximal variance, decorrelated; cumulative variance to choose; linear |
| 7. t-SNE and UMAP | Preserve local neighborhoods; distances and cluster sizes not interpretable |
| 8. Anomalies | Model the normal; IQR robust; isolation forest; contamination = business decision |
| 9. Gaussian mixtures | Probabilistic assignment, ellipsoidal shapes; EM; k by minimizing BIC |
| 10. Project | Frame the use, prepare variables, compare, characterize, document |
The threads that run through the whole course
Distance decides, so scale decides. Almost all these methods rest on distances between observations. An unstandardized wide-ranging variable imposes itself mechanically, and nothing in the result flags it — the difference from supervised learning, where a degraded score eventually raises the alarm. Standardization is not one good practice among others: it is a condition of validity.
Every method makes a shape assumption. k-means assumes comparable spheres; the Gaussian mixture, ellipses; DBSCAN, homogeneous density; PCA, straight directions. A poor result far more often comes from an unsuitable assumption than from imperfect tuning. Knowing which assumption you have just made is the real skill this course teaches.
Evaluating without ground truth requires triangulation. No single criterion suffices. You cross an internal criterion (silhouette, BIC), stability under resampling, and business usefulness — the last remaining the final arbiter. Three concordant signs beat one isolated score.
A visualization is not a proof. Module 7 is the course's guardrail: convincing projections appear even on pure noise. You explore with those images; you conclude with measurements.
The final exam
The exam includes 40 questions covering the ten modules: the nature of unsupervised tasks, the mechanics and limits of each algorithm, choosing the number of clusters, dimensionality reduction, reading projections, anomaly detection and project workflow. Several questions present mini-situations — a clustering whose silhouette scores all plateau, a projection you are asked to interpret, a dataset with unequal densities — because the course's goal is judgment, not recitation.
On success, your certificate of completion is issued immediately; its number is verifiable by any third party on the platform.
Reread the table above, then ask yourself for each row: "what assumption does this method make, and what happens when it is false?". If the answer comes — why k-means fails on two crescents, why inertia cannot choose , why you do not cluster on t-SNE coordinates — you are ready. Good luck!
Final exam
Ready to validate this course?
40 questions drawn at random from the course bank · passing score 70% · verifiable PDF certificate issued immediately on success.
Start the examYou need to be signed in to your InSkillML account with an active subscription. You can also start the exam from My courses.