Multiobjective Evolutionary Computation for Market Segmentation

Ying Liu

College of Business, California State University Long Beach, Long Beach, CA, U.S.A.

Keywords: Market Segmentation, Multiobjective Evolutionary Computation, Pareto Optimal.

Abstract: The market segmentation, by its computational essence, is a NP-hard multicriteria problem. Multiobjective

evolutionary algorithms are developed to optimize multiple objectives simultaneously and can generate a set

of Pareto optimal solutions. As a proven meta-heuristic technique, multiobjective evolutionary computation

is robust in handling different data types, various business constraints and different objective function forms.

The generated Pareto optimal solution set gives a holistic view of possible solutions that bring business

insights and allow big flexibility in solution selection. These features make the multiobjective evolution

computation a good fit for market segmentation problems. There are challenges in every phase in

implementation of multiobjective evolutionary computation for market segmentation.

1 INTRODUCTION

Market segmentation provides business decision

makers a useful perspective to understand and

differentiate customers by their needs and behaviors

(Dolnicar et al., 2018). Though market segmentation

has had an easy-to-understand conceptual definition

(Smith, 1956) for more than a half century, Wedel and

Kamakura (2000) observed that “the development of

market segmentation method has been partly

contingent on the availability of marketing data, the

advances of analytical techniques and the progress of

segmentation methodology.” As the size and richness

of customer data increase, academicians and

practitioners demand more efficient, robust, and

scalable techniques to meet the new challenges in

customer segmentation. The recent advances in

machines learning and distributed data processing

bring new capabilities and new methods to segment

customers.

Wedel and Kamakura (2000) classified market

segmentation methods into predictive and descriptive

methods. Descriptive methods use two or more sets

of variables to describe the customer segments while

predictive methods analyze the relationship between

a set of independent variables and one or more

dependent variables. This classification is helpful but

gives few clues of the complexity and the great

variance of segmentation methods in problem

https://orcid.org/0000-0002-3419-0733

definition and solution implementation. There is an

abundance of different market segmentation methods.

However, there is relatively little work done on the

computational issues of the market segmentation,

especially its multicriteria nature, to help people

understand existing methods and develop new

segmentation methods. This research investigates the

computational properties of the market segmentation

problem in section 2. Section 3 shows the complexity

and the multicreteria nature of the market

segmentation problem. Section 4 shows that the

multiobjective evolutionary computation is a good

candidate to solve the market segmentation problems.

The conclusions and future research directions are

discussed in Section 5.

2 THE COMPUTATIONAL VIEW

OF MARKET SEGMENTATION

2.1 Clustering Is a Subproblem

The challenges of market segmentation roots in its

computational properties. A fundamental task of

market segmentation is grouping customers based on

similarities in their needs and preferences. Clustering

is a common tool for this purpose (Punj and Stewart,

1983). Clustering could be generally defined as a set

Liu, Y.

Multiobjective Evolutionary Computation for Market Segmentation.

DOI: 10.5220/0010684400003063

In Proceedings of the 13th International Joint Conference on Computational Intelligence (IJCCI 2021), pages 149-154

ISBN: 978-989-758-534-0; ISSN: 2184-3236

149

of techniques that group entities that are similar in

measured characteristics. It aims to maximize the

homogeneity within the segment while maximizing

the heterogeneity among segments. Each segment is

a class of customers that marketer can identify, target,

and communicate with. In the early market

segmentation research, due to the limits of customer

data and the lack of computational power, clustering

was used widely market segmentation. However,

cluster analysis is a solution only good for single-

basis descriptive segmentation. A segmentation basis

is a set of variables used to describe a certain aspect

of customers. Different segmentation bases describe

different features of the customer or marketing mix

and have different levels of effectiveness regarding

the segmentation criteria. In a general model of

descriptive market, more than one segmentation bases

are used to take advantage of the benefits of each

segmentation basis. The model is called joint

descriptive market segmentation (Morwitz and

Schmittlein, 1992). From the view of segmentation

objectives, descriptive methods are optimized for

segment identifiability while predictive methods are

optimized for segment responsiveness. In predictive

segmentation, decision makers do not segment

customers for “clustering” purposes only. They want

actionable segments that will let them formulate

effective marketing campaign in an objective way. As

a result, in both descriptive and predictive market

segmentation, clustering is a subproblem of a general

model of market segmentation.

2.2 Computation Definition

Though the conceptual definition of market

segmentation is simple, the computational definitions

of market segmentation have been in many forms. A

common practice is to define the segmentation

problem according to different segmentation solution

techniques. Market segmentation was framed as a

clustering problem (Punj and Stewart, 1983) in early

research when the clustering techniques were used.

When the focus was shifted from descriptive

variables to response variables, the market

segmentation was framed as a segmentation problem

solved by solution procedures such as chi-squared

automatic interaction detector (CHAID) (Kass,

1980), classification and regression trees (CART)

(Breiman et al., 1984), and clusterwise regress (Spath,

1982). Unlike the clustering definition that aims to

maximize within-segment homogeneity, the segment

problem (Kleinberg et al., 2004) aims to maximize a

general utility function (usually not the segment

homogeneity to distinguish it from the clustering

problem) of all segments.

The identifiability, responsiveness and other

criteria of market segmentation demonstrate the

characteristics of both clustering and segmentation.

But market segmentation problems did not have a

computational multiobjective definition until

DeSarbo and Grisaffe (1998) used combinatorial

optimization approaches as the solution techniques.

Since then, many multiobjective optimization

approaches (Krieger and Green, 1996, Brusco et al.,

2003) were developed to solve the multicriteria

market segmentation problems. Though DeSarbo and

Grisaffe (1998) pointed out that there exists a set of

Pareto optimal solutions for a multiobjective problem

definition of market segmentation, those methods do

not generate the Pareto optimal solution set because

they are essentially single objective solution

techniques. Giving that evolutionary algorithm

generate good results for many multiobjective

optimization problems (Coello et al., 2002), Liu et al.,

(2010) applied a multiobjective evolutionary

algorithm to market segmentation. The algorithm

directly tackles the multiobjective segmentation

problem and generates a Pareto optimal solution set.

2.3 Discriminative vs. Generative

Methods

Assumptions about the segmentation data play an

important role in segmentation methods. Like the

clustering method classification scheme proposed by

Zhong and Ghosh (2003), market segmentation

methods can be classified into discriminative (or

distance/similarity-based) approaches and generative

(or model-based) approaches from their

computational assumption about the nature of data.

Discriminative methods calculate distances or

similarity between customers and segment customers

based on these measures. K-means, hierarchical

clustering, and Self-Organizing Map are typical

discriminative clustering methods. Generative

methods assume customers are from different

statistical models and try to find the parameters of the

corresponding models. Each type has its advantages

and disadvantages. Because of the direct optimization

of within-segment customer similarity,

discriminative-based segmentation methods are

usually efficient and intuitive. However, the results

usually are used as-is and no statistical inference

could be drawn from the results. There are several

advantages of generative methods. If the distribution

assumption of data is correct, they usually generate

better results than discriminative methods. The results

ECTA 2021 - 13th International Conference on Evolutionary Computation Theory and Applications

150

are more interpretable and enable statistic inference.

But generative methods such as finite mixture model

are computationally expensive when the number-of-

segments is big or there are many segmentation

variables (Wedel and Kamakura, 2000). Both the

discriminative and generative approaches may be

formulated as a data mining and/or optimization

problem whose solutions need to have confidence and

support, information content and unexpectedness

(Padmanabhan and Tuzhilin, 2003).

3 THE COMPUTATIONAL

ISSUES OF MARKET

SEGMENTATION

3.1 The Similarity Measures and

Clustering Process

Clustering is a subproblem of market segmentation.

Clustering by itself is vaguely defined and the

clustering process is hard and fuzzy (Jain et al., 1999).

The vagueness lies in the measurement of so-called

“homogeneity” or “similarity”. Punj and Stewart

(1983) discussed problems of determining the

appropriate similarity measure in market

segmentation. From several empirical experiments,

they found that each similarity measure has different

characteristics and different distance measures lead to

different clustering results. Skinner (1978) identified

three aspects of similarity measures: elevation,

scatter, and shape. In a rough sense, elevation could

be thought of as the mean of all attributes of a given

subject. Scatter is about deviation, while shape is

about the direction (up/down) of the data. The most

important finding of was that a specific distance or

similarity measure may not cover all aspects.

The hardness and fuzziness of clustering process

is explained by the impossibility theory of clustering

proved by Kleinberg (2002). It is intuitive to think of

three desired properties of any clustering process.

Scale-invariance property means that changing the

unit of distance measure should not change the

clustering result. Richness requires that a clustering

process should be able to generate all possible

partitions of clustering entities. Finally, consistency

is satisfied when the clustering result stays the same

when we increase the distance among clusters and

decrease distances within clusters. The impossibility

of clustering showed that there is no clustering

process that can satisfy all three properties

simultaneously. To avoid the limitation of clustering

process, Penaloza et al., (2017) developed a multi-

objective clustering algorithm that uses multiple

criteria to measure the quality of cluster cohesion.

Zhong and Ghosh (2003) proposed to classify the

clustering method into discriminative methods and

generative methods. In discriminative methods, also

called distance/similarity-based methods, the

similarity function is defined between pairs of

objects. In generative methods, also called model-

based methods, the similarity is defined indirectly

through the assumption of data distribution. Those

methods assume that the overall distribution of the

data is a mixture of probability distributions, each

being a different cluster (Fraley and Raftery 1998).

Even the probabilistic clustering methods assume

similarity measure, though in an indirect way. The

distinction of discriminative and generative methods

helps to understand the similarity measures among

clustering algorithms.

3.2 The Computational Complexity

The trend of big data and quick market response time

raise attentions to computational complexity of

market segmentation. Aloise et al. (2009) showed that

clustering is NP-hard even for the 2-cluster problem

using the very simple Euclidean distance to measure

similarity. Kleinberg et al., (1998) proved that most

optimization problems become NP-complete if they

are defined in a segmentation form. Krieger and

Green (1996) defined the market segmentation

problem as a 0-1 programming problem whose

computational complexity is NP-hard. Consequently,

the market segmentation problem, even framed as a

clustering problem, cannot be solved in polynomial

time. Existing methods either transform the problem

into an easy to solve version or apply heuristic

techniques to solve the problem.

3.3 The Multicriteria Nature

Marketing researchers realized that market

segmentation is a multicriteria problem from the very

beginning because customers in a segment should

have similar profiles (identifiability) and respond

similarly to a marketing mix (responsiveness) (Smith,

1956). For example, customers in a segment should

have similar demographic attributes such age,

educational level, location, etc. Identifiability makes

it easy to target a specific customer segment.

Responsiveness can be measured by customer

behaviors such as response rate or transactional

values of a marketing promotion. Simultaneously

clustering customers and predicting their responses to

marketing mix is a long-standing problem facing

Multiobjective Evolutionary Computation for Market Segmentation

151

marketing researchers. During the evolution of

market segmentation theories, more and more criteria

are added. In addition to the identifiability and

responsiveness, Wedel and Kamakura (2000) added

substantiality, accessibility, stability and actionability

criteria to evaluate whether a segmentation solution is

good or not. DeSarbo and DeSarbo (2009) added four

more criteria of differential behavior, feasibility,

profitability and projectability. At the general

conceptual level, clustering only addresses the

identifiability criterion (Brusco et al., 2003) while

other criteria such as responsiveness, profitability and

actionability must be addressed by augmented

methods.

3.4 Determining the

Number-of-Segments

The issue of determining the number-of-segments

appears in both predictive segmentation and joint

segmentation. In predictive segmentation, the

criterion of the predictive power is as important as the

criterion of the segment homogeneity. As the

number-of-segments increases, the within segment

homogeneity usually increases but the predictive

power may increase or decrease independent of the

within-segment homogeneity. Joint segmentation

consists of clustering on multiple segmentation bases,

in that, each can be thought of as an independent

clustering problem and an overall trade-off must be

made in selecting the “right” number-of-segments.

The multicriteria nature of market segmentation

means that determining the “right” number-of-

segments is a multicriteria decision and often

involves marketers’ domain knowledge.

Consequently, decision makers would like to see a set

of segmentation solutions that have different

numbers-of-segments. Those solutions give them

flexibility in investigating solutions and select the

most appropriate ones for a specific business

scenario.

4 MULTIOBJECTIVE

EVOLUTIONARY

COMPUTATION

4.1 Why Multiobjective Evolutionary

Computation?

The multicriteria nature of market segmentation

raises many issues that cannot be addressed

appropriately by traditional market segmentation

methods such as K-means and cluster-wise regression

because they only optimize one objective. Many

heuristic multiobjective methods have been

developed to address the multicriteria requirement of

market segmentation. These methods can be

classified in three categories: multi-stage method,

transformation method, and multiobjective method.

The multi-stage method solves one criterion at

one stage. For example, Kriger and Green (1996)

used K-means method to optimize group

identifiability in stage one and a heuristic algorithm

to optimize responsiveness of segments in stage two.

The disadvantage of the multi-stage approach is that

information found in one stage is not used by the other

stages because of the separated processing phases. It

is not efficient in the sense of information sharing.

The order of objective optimization often matters.

Furthermore, because each stage optimizes a single

objective, the result is often suboptimal regarding all

objectives.

The transformation method transforms multiple

criteria into one (Green and Krieger, 1991, Brusco et

al., 2003), therefore the problem can be solved by

many established single objective optimization

methods. However, it is often difficult, if not

impossible, to define an appropriate total utility or

weighted sum function to represent the multiple

criteria. Multiple criteria may be incommensurate.

For example, one criterion is the within-segment

homogeneity measure by within-segment sum of

variance and another criterion is the predictive power

in logistic regression measure by maximum

likelihood. Additionally, the transformation

procedure may put unnecessary limitations on the

search space. Global optimal solution could be lost in

transformation (Freitas, 2004).

Multiobjective evolutionary algorithms such as

NGSA II, SPEA2, and FAME (Santiago et al., 2019)

optimize multiple objectives simultaneously and

generate a set of Pareto optimal solutions. These

methods have some much-desired features. First,

multiple criteria can be independently defined in

terms of multiple optimization objects, computational

constraints, and decision variables. This avoids the

difficulties of combining multiple criteria. The

multiple objectives optimization can incorporate both

generative and discriminative measures. Second, a

multiobjective optimization method generates a set of

Pareto optimal solutions representing trade-offs

among multiple possibly conflicting objectives.

Third, a single run can generate solutions with

different number-of-segments. There is no upfront

need to determine the number-of-segments.

ECTA 2021 - 13th International Conference on Evolutionary Computation Theory and Applications

152

4.2 A Good Fit for Market

Segmentation

The evolutionary computation is the most widely

used meta-heuristic approach to solve multiobjective

optimization problems (Coello et al., 2002). It is a

good fit for market segmentation problems for several

reasons. First, it searches for optimal solution(s) using

a set of objectives simultaneously. This property

makes the algorithm efficient because all objectives

are directly used in optimization. Second, the

evolutionary algorithm can be used to find a set of

solutions that has the desired diversity. The results are

usually representative of the possible trade-offs that

are important for decision making.

Figure 1: A sample solution set that has different number-

of-segments.

The solutions in Figure 1 are generated from a

multiobjective genetic algorithm (MOEA) applied in

consumer data that include both demographic and

transactional attributes. The two minimum

optimization objectives are deviance of linear

regression and WCOS (with cluster omega squared)

for within cluster homogeneity. with the aid of data

visualization and analysis tools, the shape and the

parameters of the solution set give insights and

improve the decision-making process.

Additionally, evolution computation is less

susceptible to the shape of continuity of the Pareto

solution set because it does not make any assumptions

about data and problem properties. Finally, the

evolutionary algorithm is independent of objective

functions and decision variables. This is a very

attractive feature because the algorithm could be used

in a broad range of market segmentation problems in

different objective function forms (discrete or

continuous, concave or convex, single modal or

multimodal).

4.3 Computational Issues

Nonetheless, the application of multiobjective

evolutionary computation in market segmentation is

relatively new (O’Brien et al., 2020) and it brings

some challenges in all computational phases

The quality and diversity of the initial solution set

affects the effectiveness of evolutionary computation.

For market segmentation, existing clustering and

segmentation algorithms optimized for different

segmentation objectives can be used to generate the

initial solution set. There are not many theories to

guide the implementation and selection of

initialization algorithms.

The parameter setting in evolutionary algorithms

is another challenging task since the parameters

interact in highly non-linear ways (Lobo et al., 2007).

Grid search is often used but Bayesian optimization

can reach or surpass human expert-level optimization

on many algorithms such as structured SVM and

coevolutionary neural network (Snoek et al., 2012). It

is interesting to check its effectiveness in

evolutionary algorithm in market segmentation.

Selecting the most appropriate solution from a set

of Pareto optimal solutions consisting of solutions

with different number-of-segments is another

interesting research topic. The shape of the Pareto

front, the parameters of each solution, and the

practical constraints are factors to be considered.

Visualization tools and multiobjective data analysis

techniques are helpful in determining the number-of-

segments and solution selection.

5 CONCLUSIONS

There is a good match between the computational

essence of market segmentation problem and the

multiobjective evolutionary computation.

Multiobjective evolutionary computation brings a

new perspective to multicriteria market segmentation

in its computational model definition, optimization

process and the solution set analysis. It comes with

challenges in all phases of evolutionary computation.

Given the meta-heuristic nature of the multiobjective

evolutionary computation, it is a research topic to use

new algorithms in solution initialization, parameter

setting and solution selection. There is a need for

more empirical evaluation in different business

settings.

All Solutions

0.45

0.55

0.65

0.75

0.85

0.95

1.05

100 300 500 700 900 1100 1300

Devi ance

WCOS

3-Segment 4-Segment 5-Segment 6-Segment 7-Segment 8-Segment

Multiobjective Evolutionary Computation for Market Segmentation

153

REFERENCES

Aloise, D., Deshpande, A., Hansen, P. & Popat, P. (2009).

NP-hardness of Euclidean sum-of-squares clustering.

Machine Learning, 75(2), 245-248.

Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. I.

(1984). Classification and Regression Trees.

Wadsworth International Group.

Brusco, M. J., Cradit, J. D., & Tashchian, A. (2003).

Multicriterion Clusterwise Regression for Joint

Segmentation Settings: An Application to Customer

Value. Journal of Marketing Research, 40(2), 225-234.

Coello, C. A. C., Veldhuizen, D. A. V., & Lamont, G. B.

(2002). Evolutionary Algorithms for Solving Multi-

Objective Problems. Kluwer Academic Publishers.

DeSarbo, W. S., & Desarbo, C. F. (2007). A Generalized

Normative Segmentation Methodology Emply

Conjoint Analysis. Conjoint Measurement: Methods

and Applications, 4

ed, Springer Berlin.

DeSarbo, W. S., & Grisaffe, D. (1998). Combinatorial

Optimization Approaches to Constrained Market

Segmentation: An Application to Industrial Market

Segmentation. Marketing Letters, 9(2), 115-134.

Dolnicar, S., Grun, B., & Leisch, F. (2018). Market

Segmentation Analysis. Springer.

Fraley, C., & Raftery, A. E. (1998). How Many Clusters?

Which Clustering Methods? Answers Via Model-Based

Cluster Analysis. The Computer Journal, 41(8): 578-

588.

Freitas, A. (2004). A Critical Review of Multi-Objective

Optimization in Data Mining: a Position Paper.

SIGKDD Explorations, 6(2), 77-86.

Green, P. E., & Krieger, A. M. (1991). Segmenting Markets

with Conjoint Analysis. Journal of Marketing, 55(4),

20-31.

Jain, A. K., Murthy, M. N., & Flynn, P. J. (1999). Data

Clustering: A review. ACM Computing Surveys, 31(3),

264-323.

Kass, G. V. (1980). An Exploratory Technique for

Investigating Large Quantities of Categorical Data.

Applied Statistics, 29(2), 119-127.

Kleinberg, J. (2002). An Impossibility Theorem for

Clustering. Proceedings of the 15th Conference on

Neural Information Processing Systems (NIPS).

Vancouver, Canada, 463-470.

Kleinberg, J., Papadimitriou, C., & Raghavan, P. (2004).

Segmentation Problems. Journal of the ACM, 51(2),

263-280.

Krieger, A. M., & Green, P. E. (1996). Modifying Cluster-

Based Segments to Enhance Agreement with an

Exogenous Response Variable. Journal of Marketing

Research, 33, 351-363.

Liu, Y, Ram, S., Lusch, R. F., & Brusco, M. (2010).

Multicriterion Market Segmentation: A New Model,

Implementation, and Evaluation. Marketing Science,

29(5), 880-894.

Lobo, F. G., Lobo, F. J., & Lima, C. F., Michalewicz, Z.

(2007). Parameter Settings in Evolutionary Algorithms.

Springer.

Morwitz, V. G., & Schmittlein, D. W. (1992). Using

Segmentation to Improve Sales Forecasts Based on

Purchase Intent: Which Intenders Actually Busy?.

Journal of Marketing Research, 29, 391-405.

O’Brien, M., Liu, Y., Chen, H., & Lusch. R. (2020).

Gaining insight to B2B Relationships Through New

Segmentation Approaches: Not All Relationships Are

Equal. Expert Systems with Applications, 161.

Padmanabhan, B., & Tuzhilin, A. (2003). On the Use of

Optimization for Data Mining: Theoretical Interactions

and eCRM Opportunities. Management Science, 49(10),

1327-1343.

Penaloza, M. G. M., Montes, E. M., Ramirez, N. C., Mesa,

H. G. A. M., & Figueroa, H. V. R. (2017). Improve

Multi-objective clustering with automatic

determination of the number of clusters. Neural

Computing and Application, 28(8), 2255-2275.

Punj, G., & Stewart, D. W. (1983). Cluster Analysis in

Marketing Research: Review and Suggestion for

Application. Journal of Marketing Research, 20, 134-

148.

Santiago, A., Dorronsoro, B., Nebro, A. J., Durillo, J. J.,

Castillo, O., Fraire, H. J. (2019). A Novel Multi-

objective Evolutionary Algorithm with Fuzzy Logic

Based Adaptive Selection of Operators: FAME.

Information Science, 471, 233-251.

Skinner, H. A. (1978). Differentiating the Contribution of

Elevation, Scatter and Shape in Profile Similarity.

Educational and Psychological Measurement, 38, 297-

308.

Smith, W. R. (1956). Product Differentiation and Market

Segmentation as Alternative Marketing Strategies.

Journal of Marketing, 21, 3-8.

Snoek, J., Larochelle, H., & Adams, R. P. (2012). Practical

Bayesian Optimization of Machine Learning

Algorithms. Advances in Neural Information

Processing Systems. 25, 2960-2968.

Spath, H. (1982). A Fast Algorithm for Clusterwise Linear

Regression. Computing, 29, 175-181.

Wedel, M., & Kamakura, W. (2000). Market Segmentation:

Conceptual and Methodological Foundations. Kluwer

Academic Publisher.

Zhong, S., & Ghosh, J. (2003). A Unified Framework for

Model-based Clustering. Journal of Machine Learning

Research, 4, 1001-1037.

ECTA 2021 - 13th International Conference on Evolutionary Computation Theory and Applications

154