Hierarchical and density-based clustering
Two more families of clustering: agglomerative hierarchical clustering (distance matrix, linkage criteria, dendrogram) and DBSCAN (radius, core, border and noise points); choosing between them and k-means.
- 25 cards
- 6 topics
- Free
The outline, topic by topic
Each topic is weighted the way the exam weighs it.
- Concepts
- Questions
- Sort and link
- The numbers
- Scenarios
- Understand
A few cards from the deck
Agglomerative hierarchical clustering: what is it?
Show the answer
A method that starts with one group per individual and merges the two closest groups at each step, until only one is left.
No k to set in advance: you pick the number of groups by cutting the tree afterwards. On the other hand it handles an n × n matrix, too heavy for very large datasets.
On a dendrogram, you draw a horizontal line that cuts four branches. What does this cut give?
- Four groups, one per branch cut
- Three groups, one per merge above the line
- Only four individuals, the others being classed as noise
- Two groups, one on each side of the line
Show the answer
Four groups, one per branch cut
Each branch crossing the line carries a subtree, that is a group. The higher the line, the fewer groups; the lower, the more. Noise does not exist in hierarchical clustering.
In hierarchical clustering, …'s criterion merges the two groups whose union increases the within-cluster inertia the least.
Show the answer
In hierarchical clustering, Ward's criterion merges the two groups whose union increases the within-cluster inertia the least.
Ward looks for compact groups as k-means does, but builds the whole tree of merges. It assumes Euclidean distances.
Distance matrix: what is it?
Show the answer
A square, symmetric table giving the distance between every pair of individuals; its diagonal is zero.
For n individuals there are n(n − 1) / 2 distinct distances: 1,000 individuals need nearly 500,000, 100,000 individuals nearly 5 billion.