Questions 5

Dr. Roman Belavkin

BIS4435

Question 1

Below is a diagram of a self–organising map:

    
A B C D E
 YH 
H }
Z  >
 *
 
J Z
} H6] J Z  7
 6 o
S  
6 
>

] Z HJ Z S
  
J Z HJH Z S
 

J Z  J HH Z
S 

J Z 
J  HZ
 
Z H ZS

J  
 Z J
H
 H ZS



J Z

J S

H
Z
1 2 3

By looking at the diagram answer the following questions:

a) How many input nodes does this SOM have?

Answer: Three input nodes: node 1, 2 and 3.

b) How many output nodes does this SOM have?

Answer: Five output nodes: node A, B, C, D and E.

c) The input to an SOM can be represented by a point in an m–dimensional

space (or m–dimensional vector). How many dimensions are in the
space that this SOM is analysing?

Answer: The number of dimensions corresponds to the number of

input nodes. Thus, this SOM analyses points in three–dimensional
space. Each point is represented by three coordinates: x = (x1 , x2 , x3 ).

d) How many weights does each of the output nodes have?

1
BIS4435 2

Answer: The number of weights of every output node corresponds

to the number of input nodes. Thus, each output node has three
weights: w1 , w2 and w3 . The weights of each of the output nodes
fixate particular points in the input space: w = (w1 , w2 , w3 ).

e) The output nodes are organised in a lattice. How many dimensions

does the output lattice of this SOM have?

Answer: The output nodes A, B, C, D and E are all positioned along

a line. So, the output lattice is one–dimensional.

f ) How many output nodes can fire simultaneously?

Answer: Only one of the output nodes can fire at a time.

g) Is it important what value the output node sends when it fires?

Answer: No. In SOM it is only important which of the output nodes

fires. It does not matter what value.

h) Is there any limit on how many data points (input patterns) this SOM
can analyse?

Answer: There is no such limit. You can feed as much data as you
wish into an SOM. The input space can be infinite.

i) How many clusters can this SOM detect in the input data?

Answer: Because the number of output nodes is five, this SOM can-
not distinguish between more than five different clusters. However, it
is possible that there will be fewer than five clusters detected in the
data.

j) If node D is the winner, which output nodes are its immediate neigh-
bours?

Answer: The neighbours are node C and E.

Question 2

Consider the following self–organising map:

BIS4435 3


A B


J



J
F C
 
J

E D


The output layer of this map consists of six nodes, A, B, C, D, E and F, which
are organised into a two–dimensional lattice with neighbours connected by
lines.
Each of the output nodes has two inputs x1 and x2 (not shown on the
diagram). Thus, each node has two weights corresponding to these inputs:
w1 and w2 . The values of the weights for all output in the SOM nodes are
given in the table below:
Node A B C D E F
w1 -1 0 3 -2 3 4
w2 2 4 -2 -3 2 -1

For an input pattern x = (x1 , x2 ) the winner is determined using Euclidean

distance: q
kx − wk = (x1 − w1 )2 + (x2 − w2 )2

a) Calculate which of the six output nodes is the winner if the input
pattern is
x = (2, −4) ?
The answer should contain all the working.

Answer: First, we calculate the distance for each node:

q √ √
kx − wA k = (2 + 1)2 + (−4 − 2)2 = 9 + 36 = 45
q √ √
kx − wB k = (2 − 0)2 + (−4 − 4)2 = 4 + 64 = 68
q √ √
kx − wC k = (2 − 3)2 + (−4 + 2)2 = 1+4= 5
q √ √
kx − wD k = (2 + 2)2 + (−4 + 3)2 = 16 + 1 = 17
q √ √
kx − wE k = (2 − 3)2 + (−4 − 2)2 = 1 + 36 = 37
q √ √
kx − wF k = (2 − 4)2 + (−4 + 1)2 = 4 + 9 = 13
BIS4435 4

The winner is the node with the smallest√distance from x. Thus, in

this case the winner is node C (because 5 is the smallest distance
here).

b) After the winner for a given input x has been identified, the weights
of the nodes in SOM are adjusted using adaptation formula:

w0 = w + α h[x − w] ,

where w0 is the new weight vector, α is the learning rate, h is the

neighbourhood function. Let α = 0.5 and the neighbourhood be de-
fined as:


 1 if the node is the winner
h= 0.5 if the node is immediate neighbour of the winner
 0 otherwise


Adjust the weights in the SOM.

Answer: First, let us adapt the winner (node C). The neighbour-
hood h = 1.
! ! !
2 3 −1
x − wC = − =
−4 −2 −2
! !
−1 −0.5
α h[x − wC ] = 0.5 · 1 · =
−2 −1
! ! !
0 3 −0.5 2.5
wC = wC + α h[x − wC ] = + =
−2 −1 −3
The immediate neighbours of node C are nodes B and D. The neigh-
bourhood h = 0.5. The new weights of node B and D are:
! " ! !# !
0 0 2 0 0.5
wB = + 0.5 · 0.5 · − =
4 −4 4 2
! " ! !# !
0 −2 2 −2 −1
wD = + 0.5 · 0.5 · − =
−3 −4 −3 −3.25
All other nodes in the lattice have neighbourhood h = 0. Thus, their
weights do not change.

Question 3

What are the main similarities and differences between feed–forward neural
networks and self–organising maps?
BIS4435 5

Answer: Similarities are:

• Both are feed–forward networks (no loops).

• Nodes have weights corresponding to each link.

• Both networks require training.

The main differences are:
• Self–organising maps (SOM) use just a single output layer, they do
not have hidden layers.

• In feed–forward neural networks (FFNN) we have to calculate weighted

sums of the nodes. There are no such calculations in SOM, weights
are only compared with the input patterns using Euclidean distance.

• In FFNN the output values of nodes are important, and they are
defined by the activation functions. In SOM nodes do not have any
activation functions, and the output values are not important.

• In FFNN all the output nodes can fire, while in SOM only one.

• The output of FFNN can be a complex pattern consisting of the values

of all the output nodes. In SOM we only need to know which of the
output nodes is the winner.

• Training of FFNN usually employs supervised learning algorithms,

which require a training set. SOM use unsupervised learning algo-
rithm. There are, however, unsupervised training methods for FFNN
as well.

Question 4

Suppose that the SOM, shown in Question 1, is used to classify types of

airplanes based on three parameters: Size, speed and passenger load. The
weights of the output nodes are shown in the table below:
Node→ A B C D E
w1 3 5 1 2 5
w2 2 1 5 3 2
w3 5 1 1 2 5
Each of the three parameters is assessed on a scale from 1 to 5. For example,
small airplanes have size 1, while huge planes would have value 5. Each plane
is represented as a three–dimensional vector with coordinates corresponding
to these three parameters. Answer each of the following questions justifying
your answers:
a) How many types of planes can this SOM classify?
BIS4435 6

Answer: Maximum five types because there are only five output
nodes.

b) Which node will be the winner, if a vector representing a fighter jet is

fed into the input?

Answer: Fighter hets usually are small, very fast and carry very few
passengers. Thus, it can be represented by a vector (1, 5, 1), which is
exactly like the weights of node C (1 mark). Node C will be the
winner if the SOM is fed with a fighter jet vector, hence it must be
representing fighter jets (1 mark).

c) Which of the output nodes can represent a jumbo passanger jet?

Answer: Node E. Indeed, the weights of node D are wA = (5, 2, 5)

which corresponds to big size, not very fast speed and a heavy passen-
ger load.

d) Suppose you were asked to change the design of the SOM in order to
take into account two additional parameters: Price and fuel consump-
tion. What would you need to change in this SOM?

Answer: The planes will be described by five parameters instead of

three, so they will be represented by five–dimensional vectors. Thus,
it will be necessary to increase the number of dimensions in the input
space to five, and each node in the output layer of the SOM should
have five weights.

e) Is it possible to classify ships using this SOM? Explain your answer.

Answer: Yes and no. Yes because ships can be described in terms of
size, speed, passenger load, price and fuel consumption. However, this
classification will not take into account some other features of ships
and airplanes that put them into different classes. Therefore, in order
to make this classification more precise, one would have to increase
the number of parameters (dimensions).

Question 5

An insurance company with several thousands of customers has decided to

analyse its customers in order to unsertand better why they buy the policy.
The company collected data about its customers for the last 2 years. Each
customer’s profile was stored electronically in a database and fed into a
data warehouse, where it was assessed on 50 parameters. Discuss in a form
BIS4435 7

of essay how a self–organising map (SOM) could be used for this analysis.
Why would the results, produced by an SOM, be particularly useful for the
reports presented to strategic managers?

Answer:
• Based on description above, the centralised data warehouse contains
information about thousands of customers, each avaluated on 50 pa-
rameters. Thus, the input dataset has several thousands points in
50–dimensional space.
• The SOM would need to have 50 inputs, and the output lattice would
have arbitrary number of nodes, but probably high enough to distin-
guish between many groups of customers.
• The SOM can show the distribution of customers and their clusters
on this output map.
• Because SOM shows results on one or two dimensional feature map,
they can easily be included into a report and help to visualise the
results of the analysis.
• The individual groups of customers, discovered by SOM, can then be
further investigated and new products and promotions can be designed
specifically for these groups.

• Data analyses considering long periods of time (i.e. 2 years in this

case) and decisions on new products and promotions are more likely
to be the responsibility strategic management.

Question 6

What are the main features of an SOM that can help to analyse the business
data in a data warehouse?

Answer:
• Reduction of dimensions: The multidimensional data from a data
warehouse can be fed into an SOM, which then can display the re-
sults on a one or a two–dimensional feature map.

• Ordered display: The feature map created preserves some features of

the input space topology, such that distances between points in the
lattice (on the map) are proportional to the distances in the input
space (the data space that is analysed).

• SOM can detect clusters in data (similar data points), which are dif-
ficult to spot in huge databases using other means.
BIS4435 8

• The training algorithm in SOM is unsupervised, which is particularly

useful when not much prior information is known about the data.

• SOM can handle missing information. This is useful to classify new

data which may be incomplete.

Question 7

In which phase of decision making would a self–organising map be most

useful? Explain your answer referring to important features of SOMs.

Answer: SOMs are widely used for data analysis (data mining), and for
this reason it is most obvious to use SOM during the intelligence phase.
SOMs can help to classify data by clusters, provide convenient representation
and visualisation of the clusters, reduce number of dimensions. SOM uses
an unsupervised learning algorithm that helps to find some structure in
data without prior knowledge about its. Can handle missing data.