Session 6.3

Question 1

K-means clustering

Answer 1

• Popular method
• Centroid-based
• Find prototype data point for each cluster

Question 2

K-means clustering

Prototype-based clustering

Answer 2

• Each cluster represented by a prototype
• Other names: centroid clustering, centre-based clustering
• Example: customer segmentation. Each segment has a prototype customer and
customers similar to him/her are associated with that cluster.

Centroid - real/imaginary data point with mean characteristics of all the data points
within the cluster

Question 3

K-means clustering - how to?

Answer 3

1. Select proximity measure and specify the number of clusters (k).
2. Initiate the process by selecting centroids.
3. Assign the data points to the “nearest” centroid to form a cluster.
4. Calculate the new centroid.
5. Iterate over steps 3 and 4 until the stopping criteria are fulfilled.

Question 4

Why use k-means?

Answer 4

Strengths:

• Simple
• Efficient

Question 5

K-means weaknesses

Answer 5

• the value of k - how to determine it?
• converges to locally optimal solution
• globular/spherical clusters

Question 6

Hierarchical clustering

Answer 6

• Creates a collection of ways to group the points.

* Output: a hierarchy of potential clusterings - dendrogram.

Question 7

Why use hierarchical clustering?

Answer 7

Strengths:
• Clusters can be of any size and shape.
• Does not require to prespecify the number of clusters.

Question 8

Hierarchical clustering weaknesses

Answer 8

• Still need to decide where to split.

* Computationally inefficient.