Neural Networks in Market Segmentation

1
Neural Networks in Market Segmentation

• Jukka Hirvonen

2
Basic concepts

• Market segmentation
• A process which aims to divide a heterogeneous
  market into n distinctive and internally
  homogenous subsets, segments
• The purpose is to find segments that are
  interpretable and respond similarly to the given
  marketing strategy
• The whole concept derives from idea that clear
  market segments do exist
• Typical method Cluster analysis (K-means)

3

• Artificial neural networks
• Can be divided into two classes
• Supervised networks
• While training a supervised net, one needs to
  know the class labels of the input tubles
• In market segmentation studies, it has been used
  to classify potential customers into good and
  bad (or similar) customers
• Unsupervised networks
• Class label of each training tuple is not known
• In market segmentation studies, it is typically
  used to divide markets into n internally similar
  segments
• Examples FSCL, Hopfield network, SOM

4
First studies with supervised networks

• Fish, Barnes, and Aiken (1995)
• Purpose Classification of customers into
  pre-defined groups (segments) using supervised
  neural network method vs. discriminant and logit
  analysis
• Results Neural networks result in better
  accuracy and is better suited for classification
  due to non-existent restrecting assumptions

5
The rise of unsupervised methods

• Unsupervised methods, which are a natural choice
  for clustering purposes, started to become
  popular during the late 1990s
• However, there were still a few more articles
  where backpropagation method was used

6

• Balakrishnan, Cooper, Jacob, and Lewis (1996)
  were first to use unsupervised algorithm for
  market segmentation
• Purpose Compare FSCL neural network to
  traditional K-means clustering algorithm
• Criteria used (1) segment size, (2) segment
  means, and (3) managerial interpretation of the
  segments
• Results
• K-means better with artificial data, however
  combination of FSCL and K-means gave the most
  easily interpretable result with real-world data

7
Self-organizing maps

• Vellido, Lisboa and Meehan (1999) studied SOM as
  a visualization tool
• In addition they tested SOM alone and evaluated
  the result based on the interpretability of the
  segments
• Results
• SOM is a useful help in interpretation giving
  low-dimensional visual image of the data
• SOM was able to form segments that were easily
  interpretable

8
SOM combinations

• Kuo, Ho, and Hu (2002)
• Took the main result from Balakrishnan et al. but
  changed FSCL to SOM, hence forming a new proposed
  segmentation method SOM K-means
• Compared the classification results to
  conventional combination of Wards method
  K-means
• Evaluation was based on the Wilks lambda (ratio
  of within-group variance to total variance) and
  ability of discriminant analysis to correctly
  classify the customers into the resulting clusters

9

• Results
• With real-world data new model worked better than
  conventional one
• However, the difference between the new model and
  plain SOM was nearly non-existent

10
Kuo continues

• Kuo continued improving his model by introducing
  an improved version of K-means genetic K-means,
  which combines genetic algorithm to K-means
  algorithm
• Two articles, first one in 2004 with Chang and
  Chien, the second in 2006 with An, Wang, and
  Chung
• Both practically identical except for changes in
  words and sentence structure and minor changes in
  the proposed genetic K-means algorithm the
  former, however, was not mentioned in the latter
  as a reference

11

• The new SOM genetic K-means algorithm was
  compared to the SOM K-means algorithm presented
  in 2002
• They state that the new model outperformed the
  earlier one while using the within-group variance
  as a criterion, however, the result was not
  statistically significant
• In addition, the plain SOM, that was practically
  as good as the combination of SOM K-means was
  not tested this time

12
Back to plain SOM

• Kiang, Hu, and Fisher (2006) concentrated on
  comparing the segmentation results from extended
  SOM (i.e. SOM with given number of clusters) to
  traditional factor-analysis K-means combination
• Results
• Both methods created segments with similar
  managerial interpretation
• However, the individuals within the segments were
  different in 48 of the cases

13
Deciding the number of segments

• It is not enough to choose a segmentation model
  that minimizes the within-group variance
• One has to be able to decide the correct number
  of segments
• One may try to estimate the number of segments by
  using several different algorithms

14

• Boone and Roehm (2002)
• Purpose
• To compare two often used stopping rule
  algorithms (1) pseudo-F statistic and (2) cubic
  clusering criterion (CCC) to new challenger,
  membership clustering criterion (MCC)
• Results
• MCC combined with Hopfield (unsupervised) neural
  network gave the best accuracy rate and had no
  bias. In addition, its results were easiest to
  interpret.

15
Conclusions

• Purpose was to create a picture of the usage of
  neural networks in market segmentation from the
  first supervised models through combined SOM
  models to present plain SOM solutions
• In addition, the neural network solution for
  deciding the correct number of segments within a
  market was presented

16

• Articles with more practical purposes or articles
  that concentrate purely on neural network methods
  in clustering were not covered

17

• As a conclusion, neural network models meant for
  market segmentation are already very accurate
  WHEN they are tested with artificial data sets,
  however suggestions for future research include
  e.g.
• Creation of better ways to simulate more
  real-world-like artificial data sets
• Validation of present models with various
  different real-world data sets
• Comparisons between SOM and other unsupervised
  methods

18
Epilogue

• Especially the results from King et al. (2006)
  make me wonder, whether market segmentation as a
  marketing concept has already come to its end.
• If there are no actual, distinct, homogenous
  segments within the market, it does not make any
  difference, how perfect the segmentation
  algorithm is.

19
Questions? No? Cheers!