Lecture 8 - Unsupervised Learning I

Question 1

What do patterns tell?

Answer 1

Patterns describe or summarize the data set or parts of it

Question 2

What is cluster analysis?

Answer 2

Identifying groups of “similar” data objects?

Question 3

What are association rules?

Answer 3

Finding associations between attributes or typical combinations, like demand = high, supply = low then price = high

Question 4

What is deviation analysis?

Answer 4

Finding groups that deviate from the rest of the data, like men under 30 differ from the whole dataset.

Question 5

What is hierarchical clustering?

Answer 5

Hierarchical clustering builds clusters step by step

Question 6

What is agglomerative hierarchical clustering?

Answer 6

A bottom up strategy by first considering each data object as a separate cluster and then step by step joining clusters together

Question 7

What is divisive hierarchical clustering?

Answer 7

Starting with the whole data set as a one cluster and then dividing it to smaller ones. Seldom used, because first step has 2^(n-1) steps

Question 8

What is isotrophic distance?

Answer 8

Distance grows equally fast in all directions (like euclidean)

Question 9

What is nonisotrophic distance?

Answer 9

Distances have different weightings for different directions

Question 10

What are the b, n and x in the dissimilarity measures?

Answer 10

b = hold in both records, n = do not hold in both records, j = hold in only one of both records

Question 11

What is single linkage?

Answer 11

Dissimilarity between the two most similar data objects (so two closest ones connected)

Question 12

What is a complete linkage?

Answer 12

Dissimilarity between two most dissimilar data objects (so two that are furthest away connect)

Question 13

What is average linkage?

Answer 13

Average dissimilarity between two points of two clusters

Question 14

What is centroid linkage?

Answer 14

Distance between two centroids (mean value vectors)

Question 15

What are dendrograms?

Answer 15

The cluster merging process arranges data points in a binary tree, when drawing the data tuples at the bottom, draw a connection between clusters that are merged with the distance to the data points, then cut the dendogram at specific point to get the clusters for that distance

Question 16

What are some approaches to choose the clusters?

Answer 16

Simplest approach: Specify a minimum desired distance between clusters, stop merging when farther apart

Visual approach: Merge clusters until all data points are combined into one cluster, draw the dendrofram and find a good cut level (doesn’t have to be horizontal)

More sophisticated approaches: analyze the sequences, find a step where the step is a lot larger than previous step

Question 17

What is k-Means clustering?

Answer 17

k-means algorithm partitions data points into exactly k clusters, the k must be chosen in advance

The objective is to minimize the total intra-cluster variance

Question 18

How does k-means work?

Answer 18

Initialize the cluster centers randomly by selecting k data points, assign each data point to the cluster closest to it, update the centre, repeat until converges

Question 19

What’s the problem with k-means?

Answer 19

The results is fairly sensitive to the initial positions of cluster centers, so bad initialisation may lead to failing

Question 20

What is silhouette coefficient?

Answer 20

Silhouette value is a measure of how similar an object is to its own cluster compared to other clusters, ranges from -1 to +1, high value indicates that the value is well matched to its own cluster.

Question 21

What is density-based clustering? DBSCan?

Answer 21

Using numerical data it is possible to use density-based clustering like DBScan

1. Find a data point where the density is high, aka in distance x there are at least y other points
2. All the points in the x distance from the neighbourhood are considered to belong to one cluster
3. Expand the cluster until there is not at least y points in the distance x