COMP 2204 · Machine Learning: Unsupervised Learning

Led by Carl Linnaeus Simulacrum

K-means, hierarchical clustering, Apriori, and Eclat — finding structure in data that came without labels. Based on the K-means algorithm of Stuart Lloyd.

1. Module 1

   K-Means Clustering

   Led by Carl Linnaeus Simulacrum

   The question

   K-means converges to a local minimum that depends on initial placement. K-means++ spreads initial centres apart to avoid this. The Elbow Method suggests the optimal k. But is "minimise within-cluster sum of squares" what you mean by "cluster"?

   Outcome

   The student can implement K-means, apply the Elbow Method, and identify K-means' structural assumptions.

   Sub-units

   1. ○ 1.1 K-Means on the Mall Dataset Open this unit → ✓ ↻
   2. ○ 1.2 Essay: What Is a Cluster? Open this unit → ✓ ↻
   Begin the course → Print certificate of completion
2. Module 2

   Hierarchical Clustering

   Led by Carl Linnaeus Simulacrum

   The question

   Hierarchical clustering builds a full dendrogram without requiring k in advance. Ward linkage minimises variance increase at each merge. Looking at the dendrogram — what is the longest vertical line, and what does it mean about the natural number of clusters?

   Outcome

   The student can produce a dendrogram, choose k from it, and compare to K-means.

   Sub-units

   1. ○ 2.1 Hierarchical Clustering Open this unit → ✓ ↻
   Open this module → Print certificate of completion
3. Module 3

   Association Rule Learning: Apriori

   Led by Carl Linnaeus Simulacrum

   The question

   Support, confidence, lift. A rule with lift = 4 means customers who buy A and B buy C four times more often than chance. At what lift threshold does a rule become actionable — and what happens when you set minimum support too low?

   Outcome

   The student can implement Apriori, interpret the three metrics, and identify actionable rules.

   Sub-units

   1. ○ 3.1 Apriori on the Movie Dataset Open this unit → ✓ ↻
   Open this module → Print certificate of completion
4. Module 4

   Eclat and Rule Evaluation

   Led by Carl Linnaeus Simulacrum

   The question

   Eclat is faster than Apriori on some datasets but produces only support — no confidence, no lift. Does the added complexity of Apriori's rule metrics change the business recommendations?

   Outcome

   The student can implement Eclat and evaluate when Apriori's extra metrics add value.

   Sub-units

   1. ○ 4.1 Eclat Implementation Open this unit → ✓ ↻
   Open this module → Print certificate of completion
5. Module 5

   Unsupervised Learning in Practice

   Led by Carl Linnaeus Simulacrum

   The question

   High silhouette score does not mean good clusters. Domain knowledge determines whether the structure is real or an artefact. When should you use unsupervised learning — and what does a successful analysis actually look like?

   Outcome

   The student can apply cluster quality metrics and take a defended position on unsupervised learning's appropriate uses.

   Sub-units

   1. ○ 5.1 Final Essay: When to Use Unsupervised Learning Open this unit → ✓ ↻
   Open this module → Print certificate of completion