Cluster analysis

Question 1

Cluster analysis is a range of methods that determine if there are ___________________ of data

Answer 1

different groups or clusters

Question 2

Cluster analysis assumes _____________________

Answer 2

distinct grouping

Question 3

Cluster analysis is _____________. we are trying to see if there are any hidden groups or clusters, but don’t know how to ________ groups. The data does not come with a ________ label

Answer 3

‘unsupervised’; define; class

Question 4

In multivariate data sets, we can use ___________ to define similarity

Answer 4

distance
If two things are similar, they are probably the same sort of thing
If the other two things look very different, they are probably a different sort of thing

Question 5

For nearest neighbour clustering, define the distance between clusters as the ______________ between any of the objects

Answer 5

shortest distance

Question 6

centroid clustering use ____ between objects to define clusters. This time, consider ‘object of interest the ________ rather than the individual points that make up the cluster

Answer 6

distance, cluster

Question 7

centroid clustering: replace _____________ with __________ of the cluster to which they belong

Answer 7

individual; ‘centroid’

Question 8

Centroid clustering: use ______________ to define clusters, replace the ___________ with a ________________
The new object is _________________ (mean, median…)
This average is only meaningful is use __________

Answer 8

use distance^2 to define clusters
replace individual objects with a new ‘combined object’
the new object is between the original objects’ positions
this average is only meaningful when using distance^2

Question 9

Generally, would expect nearest neighbour to have ______ distances than centroid clustering, especially as the ____________ increases

Answer 9

shorter, size

Question 10

what does single linkage cluster analysis measure

Answer 10

what is the distance from an unclassified object to another object
e.g., nearest neighbour, centroid

Question 11

nearest neighbour sensitive to__________

Answer 11

outliers

the inclusion of an object which is far way greatly increases the ‘capture power’

Question 12

centroid methods do not account for _________________

Answer 12

the spread of within-cluster objects

Question 13

what does average linkage methods measure

Answer 13

what is the average distance from an unclassified object to the other objects in a cluster

Question 14

within group linkage method

Answer 14

create clusters with the smallest average linkage distance in them

Question 15

between group linkage

Answer 15

create clusters with the smallest average distance of newly formed links

Question 16

which hierarchical clustering method to use?

Answer 16

the nearest neighbour is purist’s approach: but sensitive to ‘outliers’: a cluster that manages to ‘grab’ a relatively distant object becomes a dominant cluster

centroid clustering difficult to interpret, but is less sensitive to outliers

average linkage methods even less sensitive to outliers, PASW uses between group average as its default

None can be easily evaluated probalistically

Question 17

K-means clustering specify _________________________.
Each cluster is defined by ________________
Often the preferred way of determining _______________

Answer 17

k, the number of clusters
centroid
exactly k clusters

Question 18

what is the method of k-means clustering

Answer 18

select k points randomly
repeat
  assign all points to nearest centroid
  re-compute the centroids
until no change in centroids

the initial centroids influence the result
use hierarchical clustering -cut off at the relevant level- to determine initial position

Question 19

How can you determine k

Answer 19

should have an idea beforehand
No fail safe way to derive
can use hierarchical cluster analysis
can plot the average within-cluster analysis

Question 20

Issues with clustering

Answer 20

1. k-means clusters might end up dividing the large clusters and combine the two small ones
2. nearest neighbour clusters are very sensitive to outliers, might end up putting outliers in separate groups, similar for centroid clustering
3. if original clusters are not spherical, e.g., elongated circular curve blocks might causes problems
   points at the tip of the crescent might end up in different groups since clustering only cares distances
4. with same data, using different clustering methods, might end up dividing the groups differently

Question 21

Cluster analysis is a very good way of describing patterns in complex, multivariate data
work well with globular clusters of _________ and __________

Answer 21

same size and density