Data Analytics

Algorithms

Topic 4
Sulfeeza Mohd Drus, CISB474
 Topic Outline

Unsupervised Machine Learning algorithms

Association Rules
Clustering
1. Partitioning
2. Hierarchical

Sulfeeza Mohd Drus, CISB474 Source: Baesens

 Clustering

Grouping data into small groups based on similarity

such that data in the same group (cluster) are as similar
as possible and data in different groups are as different
as possible.
Help users understand the natural grouping or
structure in a data set. Used either as a stand-alone
tool to get insight into data distribution or as a
preprocessing step for other algorithms.

Sulfeeza Mohd Drus, CISB474 Source: Wikipedia, Stefanowski (2008)

 Clustering
Example: How do we want to group these fruits?

Sulfeeza Mohd Drus, CISB474

 Clustering
a) Grouping based on colour

Sulfeeza Mohd Drus, CISB474

 Clustering
b) Grouping based on shape

Sulfeeza Mohd Drus, CISB474

 Clustering

Other examples:
Cluster customers based on their purchase histories, so
that a targeted marketing program can be developed
Cluster products based on the sets of customers who
purchased them
Cluster documents based on similar words
Cluster DNA sequences based on edit distance

Sulfeeza Mohd Drus, CISB474 Source: Wikipedia

Clustering
Good clustering method will produce high quality
clusters with:
High intra-cluster similarity
Low inter-cluster similarity

Inter-cluster Intra-cluster
distance is distance is
maximized Education minimized

Income

Age

Sulfeeza Mohd Drus, CISB474

 Clustering

The quality of a clustering result depends on

the similarity measure used
implementation of the similarity measure
The quality of a clustering method is also measured by
its ability to discover some or all of the hidden
patterns

Sulfeeza Mohd Drus, CISB474

 Clustering
Steps to perform cluster analysis:
Step 1:
Formulate the problem Decide on the clustering variable

Step 2:
Decide on the clustering procedure

Step 3:
Decide on the number of clusters

Step 4:
Validate and Interpret cluster solution

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
Steps to perform cluster analysis:
Step 1:
Formulate the problem Decide on the clustering variable

The objective of this step is:

To select variables that could provide a clear-cut
differentiation between segments/groups regarding a
specific managerial objective

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
Steps to perform cluster analysis:
Step 1:
Formulate the problem Decide on the clustering variable

Types and examples of clustering variables

General Specific
Observable Cultural, geographic, User status, usage
(directly measurable) demographic, socio- frequency, store and brand
economic loyalty

Unobservable Psychographics, values, Benefits, perceptions,

(inferred) personality, lifestyle attitudes, intentions,
preferences
Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011
 Clustering
Steps to perform cluster analysis:
Step 2:
Decide on the clustering procedure

Option 1: Option 2:
Hierarchical methods Partitioning methods

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering

Hierarchical methods

Agglomerative Divisive

Each object starts with their All objects start in the same
own separate cluster. Then, cluster, and gradually split up
two closest (most similar) until each object becomes in
cluster is combined and individual cluster
repeatedly performed until all
objects become in one cluster

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
Steps in hierarchical methods:
1. Measure of similarity/dissimilarity
Euclidean distance
Manhattan distance Point A

Point B
Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011
 Clustering
Lets do some exercises:
Calculate the Euclidean distance for the following data sets:
A1 = (2,10)
A2 = (2,5)
A3 = (8,4)
A4 = (5,8)
A5 = (7,5)
A6 = (6,4)
A7 = (1,2)
A8 = (4,9)

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering

To calculate Euclidean distance from A1 to A2:

d(A1, A2) = (2-2)2 + (10-5)2

= 02 + 52

= 25

=5

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
Euclidean distance matrix
A1 A2 A3 A4 A5 A6 A7 A8
A1 0 5 ? ? ? ? ? ?
A2
A3
A4
A5
A6
A7
A8

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
Euclidean distance matrix
A1 A2 A3 A4 A5 A6 A7 A8
A1 0 5 6 3.6 7.1 7.2 8.1 2.2
A2 0 6.1 4.2 5 4.1 3.1 4.5
A3 0 5 1.4 1.4 7.2 6.4
A4 0 3.6 4.1 7.2 1.4
A5 0 1.4 6.7 5
A6 0 5.4 5.4
A7 0 7.6
A8 0

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
2. Choose clustering algorithm
Single linkage
The distance between two objects is defined to be the smallest distance
possible between them.
If both objects are clusters, the distance between the two closest members are
used.

Complete linkage
This method is much like the single linkage, but instead of using the
minimum of the distances, we use the maximum distance

Average linkage
The distance between two clusters is defined as the average distance between
all pairs of the two clusters members

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
2. Choose clustering algorithm

Sulfeeza Mohd Drus, CISB474

 Clustering
2. Perform hierarchical clustering using single link algorithms for
the sample datasets
A1 A2 A3 A4 A5 A6 A7 A8
A1 0 5 6 3.6 7.1 7.2 8.1 2.2
A2 0 6.1 4.2 5 4.1 3.1 4.5
A3 0 5 1.4 1.4 7.2 6.4
A4 0 3.6 4.1 7.2 1.4
A5 0 1.4 6.7 5
A6 0 5.4 5.4
A7 0 7.6
A8 0

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
Initial state 8 clusters
A1 A2 A3 A4 A5 A6 A7 A8
A1 0 5 6 3.6 7.1 7.2 8.1 2.2

A2 0 6.1 4.2 5 4.1 3.1 4.5

A3 0 5 1.4 1.4 7.2 6.4

A4 0 3.6 4.1 7.2 1.4

A5 0 1.4 6.7 5

A6 0 5.4 5.4

A7 0 7.6

A8 0

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering A1 A2 A3 A4, A5, A7
A8 A6
A1 0 5 6 2.2 7.1 8.1

1st clustering cycle A2 0 5 4.2 4.1 3.1

A3 0 5 1.4 7.2

A4, A8 0 3.6 7.2

A5, A6 0 5.4

A7 0

5.0

4.0

3.0

2.0

1.0

A4 A8 A5 A6 A3

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
A1 A2 A4, A8 A3, A7
A5, A6

2nd clustering cycle A1

A2
0 5
0
2.2
4.2
6
4.1
8.1
3.1

A4, A8 0 3.6 7.2

A3, A5, A6 0 5.4

A7 0

5.0

4.0

3.0

2.0

1.0

A4 A8 A5 A6 A3 A1

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
A2 A1, A4, A3, A5, A7
A8 A6

3rd clustering cycle A2

A1, A4, A8
0 4.2
0
4.1
3.6
3.1
7.2

A3, A5, A6 0 5.4

A7 0

5.0

4.0

3.0

2.0

1.0

A4 A8 A5 A6 A3 A1 A2 A7

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
A2, A7 A1, A4, A3, A5,
A8 A6
4th clustering cycle A2, A7 0 4.2 4.1

A1, A4, A8 0 3.6

A3, A5, A6 0

5.0

4.0

3.0

2.0

1.0

A4 A8 A5 A6 A3 A1 A2 A7

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
A2, A7 A1, A4, A8,
A3, A5, A6
5th clustering cycle A2, A7 0 4.1

A1, A4, A8, A3, A5, 0

A6

5.0

4.0

3.0

2.0

1.0

A4 A8 A5 A6 A3 A1 A2 A7

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
Steps to perform cluster analysis:
Partitioning methods

The most common one is k-means

However, for partitioning methods, we need to
determine our initial cluster.
We will get back to k-means later

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
Steps to perform cluster analysis:
Step 3:
Decide on the number of clusters

The main objective is:

To achieve maximum inter-variance and minimum
intra-variance
One of the method to determine number of clusters is
by using elbow plot
Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011
 Clustering
Elbow criterion:
Choose a number of clusters so that by adding another cluster
does not add sufficient additional information

Sulfeeza Mohd Drus, CISB474 Source: Derek Kane, Venkatesan (2007)

 Clustering - revisit
Steps to perform k-means cluster analysis:
1. Choose the number of cluster, k
2. Generate k random points as cluster centroids
3. Assign each point to the nearest cluster centroid
4. Recompute the new cluster centroid
5. Repeat the two previous steps until the convergence
criterion is met

Sulfeeza Mohd Drus, CISB474 Source: Venkatesan, 2007

 Clustering - revisit
Lets say we use the same data set:
A1 A2 A3 A4 A5 A6 A7 A8
A1 0 5 6 3.6 7.1 7.2 8.1 2.2

A2 0 6.1 4.2 5 4.1 3.1 4.5

A3 0 5 1.4 1.4 7.2 6.4

A4 0 3.6 4.1 7.2 1.4

A5 0 1.4 6.7 5

A6 0 5.4 5.4

A7 0 7.6

A8 0

Sulfeeza Mohd Drus, CISB474 Source: Venkatesan, 2007

 Clustering - revisit
Steps to perform k-means cluster analysis:
1. Choose the number of cluster, k
Lets say we start with k=3
2. Generate k random points as cluster centroids
And we choose the following points as the cluster centroids
A1, A4, A7

Sulfeeza Mohd Drus, CISB474 Source: Venkatesan, 2007

 Clustering - revisit
Then, we need to calculate the distance for each point to each
centroid (seed) Distance from A1 to: Distance from A2 to:

Seed 1 0 Seed 1 5

Seed 1 Seed 2 3.6 Seed 2 4.2

Seed 3 8.1 Seed 3 3.1

Seed 2 Distance from A3 to: Distance from A4 to:

Seed 1 6 Seed 1 3.6

Seed 2 5 Seed 2 0

Seed 3 7.2 Seed 3 7.2

Distance from A5 to: Distance from A6 to:

Seed 1 7.1 Seed 1 7.2

Seed 2 3.6 Seed 2 4.1

Seed 3 6.7 Seed 3 5.4

Seed 3
Distance from A7 to: Distance from A8 to:

Seed 1 8.1 Seed 1 2.2

Seed 2 7.2 Seed 2 1.4

Seed 3 0 Seed 3 7.6

Sulfeeza Mohd Drus, CISB474
 Clustering - revisit
3. Assign each point to the nearest cluster centroid
Distance from A1 to: Distance from A2 to:

Seed 1 0 Seed 1 5

Seed 1 Seed 2 3.6 Seed 2 4.2

Seed 3 8.1 Seed 3 3.1

Seed 2 Distance from A3 to: Distance from A4 to:

Seed 1 6 Seed 1 3.6

Seed 2 5 Seed 2 0

Seed 3 7.2 Seed 3 7.2

Distance from A5 to: Distance from A6 to:

Seed 1 7.1 Seed 1 7.2

Seed 2 3.6 Seed 2 4.1

Seed 3 6.7 Seed 3 5.4

Seed 3
Distance from A7 to: Distance from A8 to:

Seed 1 8.1 Seed 1 2.2

Seed 2 7.2 Seed 2 1.4

Seed 3 0 Seed 3 7.6

Sulfeeza Mohd Drus, CISB474
 Clustering - revisit
4. Recompute the new cluster centroid

Seed 1
Seed 1 = (2,10)
Seed 2 Seed 2 = (8+5+7+6+4)/5,
Seed 2 = (4+8+5+4+9)/5
Seed 2 = (6, 6)
Seed 3 = (2+1)/2, (5+2)/2
Seed 2 = (1.5, 3.5)
Seed 3

Sulfeeza Mohd Drus, CISB474

 Clustering - revisit

Seed 1 Seed 1 = (2,10)

Seed 2 = (8+5+7+6+4)/5,
Seed 2 = (4+8+5+4+9)/5
New Seed 2 Seed 2 = (6, 6)
Seed 3 = (2+1)/2, (5+2)/2
New Seed 3 Seed 2 = (1.5, 3.5)

Sulfeeza Mohd Drus, CISB474

 Clustering - revisit

5. Repeat the two previous steps until the convergence

criterion is met
Convergence criterion when the assignment of points in clusters do
not change over multiple iterations

Sulfeeza Mohd Drus, CISB474

 Clustering
Steps to perform cluster analysis:
Step 4:
Validate and Interpret cluster solution

Stability and Validity

Stability is evaluated using different clustering procedures
on the same data and testing whether these yield the same
results.
Eg: for hierarchical clustering, use different distance
measures
Validity can be evaluated using criterion validity

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

 Clustering
Steps to perform cluster analysis:
Step 4:
Validate and Interpret cluster solution

Profiling of cluster
Interpreting the clusters by examining the cluster centroids
It helps to shed light on whether the segments are
conceptually distinguishable
This information will also help to find the meaningful label
or name for the cluster to adequately reflects the objects in
the cluster

Sulfeeza Mohd Drus, CISB474 Source: Mooi & Sarstedt, 2011

Sulfeeza Mohd Drus, CISB474