For the exclusive use of J. Greene, 2021.

UV0745
Rev. Mar. 28, 2018

Cluster Analysis for Segmentation

Introduction

We all understand that consumers are not all alike. This provides a challenge for the development and
marketing of profitable products and services. Not every offering will be right for every customer, nor will
every customer be equally responsive to your marketing efforts. Segmentation is a way of organizing
customers into groups with similar traits, product preferences, or expectations. Once segments are identified,
marketing messages and in many cases even products can be customized for each segment. The better the
segment(s) chosen for targeting by a particular organization, the more successful the organization is assumed
to be in the marketplace. Since its introduction in the late 1950s, market segmentation has become a central
concept of marketing practice.

Segments are constructed on the basis of customers’ (1) demographic characteristics, (2) psychographics,
(3) desired benefits from products/services, and (4) past-purchase and product-use behaviors. These days,
most firms possess rich information about customers’ actual purchase behavior, geodemographic, and
psychographic characteristics. In cases where firms do not have access to detailed information about each
customer, information from surveys of a representative sample of the customers can be used as the basis for
segmentation.

An Example

Consider Geico, an auto insurance company. Suppose Geico hypothetically plans to customize its auto
insurance offerings and needs to understand what its customers view as important from their insurance
provider. Geico can ask its customers to rate how important the following two attributes are to them when
considering the type of auto insurance they would use:
 savings on premium
 existence of a neighborhood agent

The importance of the attributes is measured using a seven-point Likert-type scale, where a rating of one
represents not important and seven represents very important. Unless every respondent who is surveyed gives
identical ratings, the data will contain variations that you can use to cluster or group respondents together, and
such clusters are the segments. The groupings of customers are most similar to each other if they are part of
the same segment and most different from each other if they are part of different segments. By inference,

This technical note was prepared by Rajkumar Venkatesan, Associate Professor of Business Administration. Copyright  2007 by the University of
Virginia Darden School Foundation, Charlottesville, VA. All rights reserved. To order copies, send an email to sales@dardenbusinesspublishing.com. No
part of this publication may be reproduced, stored in a retrieval system, used in a spreadsheet, or transmitted in any form or by any means—electronic, mechanical, photocopying,
recording, or otherwise—without the permission of the Darden School Foundation. Our goal is to publish materials of the highest quality, so please submit any
errata to editorial@dardenbusinesspublishing.com.

This document is authorized for use only by Jamison Greene in ND MBA Market Research, Fall 2021-Mod 2 taught by Joe Urbany, University of Notre Dame from Oct 2021 to Feb 2022.
 For the exclusive use of J. Greene, 2021.

Page 2 UV0745

then, actions taken toward customers in the same segment should lead to similar responses, and actions taken
toward customers in different segments should lead to different responses.

Another way of saying this is that the aspects of auto insurance that are important to any given customer
in one segment will also be important to other customers in that same segment. Furthermore, those aspects
that are important to that customer will be different from what is important to a customer in a different
segment. Figure 1 shows what the analysis in this example might look like:

Figure 1. Segmentation of Geico customers.

Premium Savings
Very Important

Segment A Segment C
(49%) (15%)

Agent Not Agent Very

Important Important

Segment B
(36%)

Premium
Savings Not
Important

Source: All figures created by case writer, unless otherwise noted.

The analysis shows three distinct segments. The majority of Geico’s customers (Segment A, 49%) prefer
savings on their premium, and they do not prefer having a neighborhood agent. Customers who belong to
Segment B (about 36%) prefer having a neighborhood agent and premium savings is not important to them.
Some customers (Segment C, 15%) prefer both the savings on their premium as well as a neighborhood
agent. This analysis shows that Geico can benefit by adding an offline channel (i.e., developing a network of
neighborhood agents) to serve Segment B and also charge a higher premium to them for providing this
convenience. Of course, the caveat is the increased competition with other insurance providers, such as
Allstate and State Farm, who already provide this service.

Cluster Analysis

Cluster analysis is a class of statistical techniques that can be applied to data that exhibit natural
groupings. Cluster analysis makes no distinction between dependent and independent variables. The entire set
of interdependent relationships is examined. Cluster analysis sorts through the raw data on customers and
groups them into clusters. A cluster is a group of relatively homogeneous customers. Customers who belong to
the same cluster are similar to each other. They are also dissimilar to customers outside the cluster,

This document is authorized for use only by Jamison Greene in ND MBA Market Research, Fall 2021-Mod 2 taught by Joe Urbany, University of Notre Dame from Oct 2021 to Feb 2022.
 For the exclusive use of J. Greene, 2021.

Page 3 UV0745

particularly customers in other clusters. The primary input for cluster analysis is a measure of similarity
between customers, such as correlation coefficients, distance measures, and association coefficients.

The following are the basic steps involved in cluster analysis:

1. Formulate the problem—select the variables you want to use as the basis for clustering.
2. Compute the distance between customers along the selected variables.
3. Apply the clustering procedure to the distance measures.
4. Decide on the number of clusters.
5. Map and interpret clusters—draw conclusions—illustrative techniques like perceptual maps are
useful.

Distance Measures

The main input into any cluster analysis procedure is a measure of distance between individuals who are
being clustered. The objective of a distance measure is to quantify the difference between two individuals on
the variables you are using for the segmentation. A shorter (longer) distance between two individuals would
imply they have similar (dissimilar) preferences on the segmentation variables. Distance between two
individuals is obtained through a measure called Euclidean distance. If two individuals, Joe and Sam, are being
clustered on the basis of n variables, then the Euclidean distance between Joe and Sam is represented as:

Euclidean distance =
x Joe,1  x Sam,1   ...  x Joe,n  x Sam,n 
2 2

where:

xJoe,1 = the value of Joe along variable 1, and

XSam,1 = the value of Sam along variable 1.

A pairwise distance matrix among individuals who are being clustered can be created using the Euclidean
distance measure. Extending the preceding example, consider three individuals—Joe, Sam, and Sara—who
are being clustered based on their preference for Premium Savings and a Neighborhood Agent. The
importance ratings on these two attributes for Joe, Sam, and Sara are shown in Table 1.

Table 1. Sample data for cluster analysis.

Individual Name Importance Score
Premium Savings Neighborhood Agent
Joe 4 7
Sam 3 4
Sara 5 3

This document is authorized for use only by Jamison Greene in ND MBA Market Research, Fall 2021-Mod 2 taught by Joe Urbany, University of Notre Dame from Oct 2021 to Feb 2022.
 For the exclusive use of J. Greene, 2021.

Page 4 UV0745

The Euclidean distance between Joe and Sam is obtained as:

Euclidean distance (Joe, Sam) =

4  32  7  42 = 3.2.
The first term in this Euclidean distance measure is the squared difference between Joe and Sam on the
importance score for Premium Savings, and the second term is the squared difference between them on the
importance score for Neighborhood Agent. The Euclidean distances are then computed for each pairwise
combination of the three individuals being clustered to obtain a pairwise distance matrix. The pairwise
distance matrix for Joe, Sam, and Sara is shown in Table 2.

Table 2. Pairwise distance matrix.

Joe Sam Sara
Joe 0 3.2 4.1
Sam 0 2.2
Sara 0

The distance between Joe and Sam is 3.2, as shown in Table 2. This pairwise distance matrix is then provided
as an input to a clustering algorithm.

K-Means Clustering Algorithm

K-means clustering belongs to the nonhierarchical class of clustering algorithms. It is one of the more
popular algorithms used for clustering in practice because of its simplicity and speed. It is considered to be
more robust to different types of variables, is more appropriate for large datasets that are common in
marketing, and is less sensitive to some customers who are outliers (in other words, extremely different from
others).

For K-means clustering, the user has to specify the number of clusters required before the clustering
algorithm is started. The basic algorithm for K-means clustering is as follows:

Algorithm

1. Choose the number of clusters, 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 some convergence criterion is met. Usually the convergence
criterion is that the assignment of customers to clusters has not changed over multiple iterations.

A cluster centroid is simply the average of all the points in that cluster. Its coordinates are the arithmetic
mean for each dimension separately over all the points in the cluster. Consider Joe, Sam, and Sara in the
previous example. Let’s represent them based on their importance ratings on Premium Savings and
Neighborhood Agent as: Joe = {4,7}, Sam = {3,4}, Sara = {5,3}. If you assume that they belong to the same
cluster, then the center for their cluster is obtained as:

Cluster centroid Z = (z1,z2) = {(4+3+5)/3, (7+4+3)/3}.

This document is authorized for use only by Jamison Greene in ND MBA Market Research, Fall 2021-Mod 2 taught by Joe Urbany, University of Notre Dame from Oct 2021 to Feb 2022.
 For the exclusive use of J. Greene, 2021.

Page 5 UV0745

z1 is measured as the average of the ratings of Joe, Sam, and Sara on Premium Savings. Similarly, z2 is
measured as the average of their ratings on Neighborhood Agent. Figure 2 provides a visual representation
of K-means clustering.

Figure 2. Visual representation of K-means clustering.

Number of clusters

One of the main issues with K-means clustering is that it does not provide an estimate of the number of
clusters that exists in the data. The K-means clustering has to be repeated several times with different “Ks”
(or number of clusters) to determine the number of clusters that is appropriate for the data. A commonly
used method to determine the number of clusters is the elbow criterion.

The elbow criterion states that you should choose a number of clusters so that adding another cluster
does not add sufficient information. The elbow is identified by plotting the ratio of the within cluster variance to
between cluster variance against the number of clusters. The within cluster variance is an estimate of the average
of the variance in the variables used as a basis for segmentation (Importance Score ratings for Premium
Savings and Neighborhood Agent in the Geico example) among customers who belong to a particular cluster.
The between cluster variance is an estimate of the variance of the segmentation basis variables between
customers who belong to different segments. The objective of cluster analysis (as mentioned before) is to
minimize the within cluster variance and maximize the between cluster variance. Therefore, as the number of
clusters is increasing, the ratio of the within cluster variance to the between cluster variance will keep
decreasing.

But at some point, the marginal gain from adding an additional cluster will drop, giving an angle in the
graph (the elbow). In Figure 3, the elbow is indicated by the circle. The number of clusters chosen should
therefore be 3.

This document is authorized for use only by Jamison Greene in ND MBA Market Research, Fall 2021-Mod 2 taught by Joe Urbany, University of Notre Dame from Oct 2021 to Feb 2022.
 For the exclusive use of J. Greene, 2021.

Page 6 UV0745

Figure 3. Elbow plot for determining number of clusters.

Elbow Plot

300
Ratio of Within Cluster to Between Cluster

250

200
Variance

150

100

50

0
1 2 3 4 5 6 7
Num ber of Clusters

It should also be noted that the initial assignment of cluster seeds has a bearing on the final model
performance. Some common methods for ensuring the stability of the results obtained from K-means
clustering include:
 Running the algorithm multiple times with different starting values. When using random starting
points, running the algorithm multiple times will ensure a different starting point each time.
 Splitting the data randomly into two halves and running the cluster analysis separately on each half.
The results are robust and stable if the number of clusters and the size of different clusters are similar
in both halves.

Profiling Clusters

Once clusters are identified, the description of the clusters in terms of the variables used for clustering—
or using additional data such as demographics—helps to customize marketing strategy for each segment. This
process of describing the clusters is called profiling. Figure 1 is an example of such a process. A good deal of
cluster-analysis software also provides information on which cluster a customer belongs to. This information
can be used to calculate the means of the profiling variables for each cluster. In the Geico example, it is useful
to investigate whether the segments also differ with respect to demographic variables such as age and income.
In Table 3, consider the distribution of age and income for Segments A, B, and C as provided in Figure 1.

This document is authorized for use only by Jamison Greene in ND MBA Market Research, Fall 2021-Mod 2 taught by Joe Urbany, University of Notre Dame from Oct 2021 to Feb 2022.
 For the exclusive use of J. Greene, 2021.

Page 7 UV0745

Table 3. Age and income distribution for segments.

Segment Mean Range

Age Income ($) Age Income ($)
A 21 15,000 16–25 0–25,000
B 45 120,000 33–55 75,000–215,000
C 39 40,000 39–54 24,000–60,000

Mean represents the averages of age and income of customers belonging to a particular segment. Range
represents the minimum and maximum values of age and income for customers in a segment. Whereas the
mean is useful for identifying the central tendency of a segment, the range helps in evaluating whether the
segments overlap with regards to the profile variable.

From Table 3, you see that Segment A customers who prefer high savings on their premium and do not
prefer having a neighborhood agent tend to be younger and have low income. These could probably be
college students or recent graduates who are more comfortable with transacting online. Customers who
belong to Segment B, on the other hand, are older and have higher income levels. It would be interesting to
evaluate if these customers also tend to be married with kids. The security of having a neighborhood agent
who can help in case of an accident or emergency is very important to them, and they do not mind paying a
higher price for this sense of security. These customers may also not be comfortable in transacting (or
providing personal information) online.

Finally, while Segment C customers are as old as Segment B customers, they tend to have lower incomes
and do not prefer to have a neighborhood agent (probably because of low disposable incomes). Identification
of the segments through these demographic characteristics enables a marketer to target as well as customize
communications to each segment. For example, if Geico decides to develop a network of neighborhood
agents, it can first focus on neighborhoods (identified through their zip codes) that match the profile of
Segment B customers.

Conclusion

Given a segmentation basis, the K-means clustering algorithm would identify clusters and the customers
that belong to each cluster. The management, however, has to carefully select the variables to use for
segmentation. Criteria frequently used for evaluating the effectiveness of a segmentation scheme include:
identifiability, sustainability, accessibility, and actionability.1 Identifiability refers to the extent that managers can
recognize segments in the marketplace. In the Geico example, the profiling of customers allows you to
identify customer segments through their age and income information. PRIZM and ACORN are popular
databases that provide geodemographic information that can be used for segmentation as well as profiling.
The sustainability criterion is satisfied if the segments represent a large enough portion of the market to ensure
profitable customization of the marketing program. The extent to which managers can reach the identified
segments through their marketing campaigns is captured by the accessibility criterion. Finally, actionability refers
to whether customers in the segment and the marketing mix necessary to satisfy their needs are consistent
with the goals and core competencies of the firm. The success of any segmentation process therefore requires
managerial intuition and careful judgment.

1 For more details, refer to Wagner Kamakura and Michel Wedel, Market Segmentation: Conceptual and Methodological Foundations, 2nd ed. (Norwell, MA:

Kluwer Academic Publishers, 2000).

This document is authorized for use only by Jamison Greene in ND MBA Market Research, Fall 2021-Mod 2 taught by Joe Urbany, University of Notre Dame from Oct 2021 to Feb 2022.