4b. Continued.

Question 1

What is this clustering called?

Answer 1

Hierarhical clustering

Because:
It starts with each object as its own cluster at the bottom, then merges clusters step-by-step.

Question 2

How useful are step 1 and 5?

Answer 2

NOT useful at ALL.
(you have to figure out where you stop, step 2, 3 or 4?)

Question 3

How do we calculate the similarity between two observations?
(How do we group them together?)

Answer 3

CALCULATE THE DISTANCE BETWEEN OBJECTS!

Question 4

What are the 3 main formulas used to calculate the distances between objects?

Answer 4

1. Euclidean
2. City-block
3. Chebychev

Question 5

How does the “Euclidean” distance-calculation work?

Answer 5

Measure the direct distance from point B to C

Question 6

How does the “City-Block” distance-calculation work?

Answer 6

City block calculates distances similar to city blocks in like new york (so its difference along x axis + difference along y axis)

Question 7

How does the “Chebychev” distance-calculation work?

Answer 7

Chebychev distance looks at both distances (of city block), and takes biggest one only, so in this case the horizontal is bigger (x diff is larger), so thats the chebychev distance.

Question 8

Which of the 3 distance-calculation formulas is most important/ popular?

Answer 8

Euclidean (direct distance from point B to C)

Question 9

What are the 5 “Clustering Algorithms”?

Answer 9

1. Single Linkage
2. Complete Linkage
3. Average Linkage
4. Centroid
5. Ward’s Method

Question 10

What is the difference between “Single” and “Complete” Linkage?

Answer 10

They are OPPOSITES!

single = groups based on shortest distance between any two members in the clusters (nearest neighbour)

complete = based on longest distance between any two members (furthest neighbour)

Question 11

What is the problem with “Single Linkage”, and why is it mainly used?

Answer 11

Its a CONTRACTIVE method
It tends to form few BIG groups (chains)
Lecturer says it still serves a useful purpose though:
–> QUICKLY spotting any outliers

Question 12

What is “Average Linkage”?

Answer 12

Distance between two clusters is defined as the average distance between all pairs of the two clusters members

Question 13

What is “Centroid”?

Answer 13

The geometric center (centroid) of each cluster is computed first, then that distance is considered

Question 14

Why can “Average Linkage” and “Centroid” be useful “Clustering Algorithms”?

Answer 14

Because they are CONSERVATIVE methods which produce HOMOGENOUS CLUSTERS of SIMILAR SIZES

Question 15

What is “Ward’s Method”?

Answer 15

The object whose merger increases the overall within-cluster variance to the smallest possible degree, are combined

(so rather than looking at “closest by distance” in the usual sense, it looks at least added variance/SSE)

Question 16

Which “Clustering Algorithm” is superior?

(if clusters can be expected to have similar sizes and data does NOT include outliers)

Answer 16

Ward’s Method

Question 17

What is this called?

Answer 17

Dendogram

Question 18

From the dendogram, how do you determine the number of clusters?

Answer 18

You read dendogram from left to right
- find the longest horizontal line (this is where you “CUT”)

Question 19

Is the dendogram always give you the exact answer for the number of clusters?

Answer 19

NOT NECESSARILY

Question 20

After you read from the dendogram, what is the most important factor to determine the number of clusters further?

Answer 20

INTERPRETABILITY + VALIDITY

The clusters must be manageable and large enough to warrant attention

Question 21

How can you assess “STABILITY”?

Answer 21

Stability:
- cut dataset in two, and test whether the same test yields the same results)

Question 22

How can you assess “VALIDITY”?

Answer 22

As in… If clusters are hard to describe or you can’t identify who’s in them in practice → low validity, even if the algorithm produced them (aka..USELESS)