Since about 1995, there have been some
interesting new developments in market segmentation research. The following
sections discuss those recent advancements and include:
- Multidimensional segmentation,
- Artificial neural networks,
- Latent class models,
- Fuzzy and overlapping clustering,
and
- Occasion-based segmentation.
Multidimensional Segmentation
There is no reason to limit the basis for segmentation to only one type
of variable when many criteria actually determine buyers' responses
to the selling proposition.
A segmentation scheme based on multiple dimensions, using separate segmentation
schemes for each one, is often more useful and more flexible for planning
marketing strategy and executing marketing tactics.
Thus, marketers may consider different segmentations on a sample of
buyers using different bases such as: stated needs, benefits, and amount
spent in the category.
In the past, such segmentation schemes were deemed too confusing and
produced too many segments for marketing managers to address effectively.
Yet, in an era of micro-niche marketing and direct marketing tools,
many market planners now consider market segmentation schemes that support
finer targeting efforts.
Artificial Neural Networks
Starting in the early 1990s, ANNs have been developed to address host
of analytical problems. Both the appeal and the bane of ANNs is the
they do not require any particular underlying model formulation or data
structure, as do regression analysis, logit modeling, or factor analysis.
In general, ANNs are given a set of input variables and a set of known
outcomes, and the algorithm is asked to find the best relationship between
the inputs and the outputs. It does this by initially forming a trial
relationship on a subset of the data, called the learning set.
The algorithm then backs up through one or more "hidden layers"
of input junctures, or neurons, and adjusts the weight of each input
to that neuron to maximize its contribution to accurately predicting
the outcome. Results are validated with a third sample, the validation
sample
There are some specialized neural networks designed to cluster cases
of data. These fall in the class of unsupervised neural network meaning
that the outcomes are not prespecified.
One of the best known of these clustering ANNs is the Kohonan Self-Organizing
Map. All ANNs of this type require a large number of cases because they
need a large leaning sample, a large test sample, and a large validation
sample.
The usefulness of the clustering solution seems dependent on the initial
selection of seeds or the shape of the transform function. Many alternative
runs may be necessary to find an acceptable solution.
Another issue with ANNs is that they can overlearn. Determining when
to stop an ANN from learning is a problem that has not yet been fully
solved.
Latent Class Models (Mixture Models)
Basically, latent class models (LCMs) enable the user to simultaneously
optimize a research function and find clusters of cases within that
framework. In general, the model may be applied to almost any dependency
model – such as regression, logit and discriminant.
Software is being rapidly developed to apply LCM to a variety of standard
optimization models.
The problem with defining market segments using any of the dependency
methods, including LCM and CHAID, is that you are assuming the market
is segmented based on optimizing the explained variance in a single
dependent variable. This is seldom sufficient for strategic and many
tactical market segmentation efforts. However, the methods can be very
useful for better understanding market structures.
Fuzzy and Overlapping Clustering
Most clustering algorithms are programmed so that all cases are assigned
to one and only one cluster. The basic idea in fuzzy (or overlapping)
clustering is to allow a single case to be assigned to more than one
cluster. Currently, there is no widely available software to handle
this procedure, and there may be little need for it.
Think about a situation where you ask respondents to complete a conjoint
trade-off task about their beer selection preferences in different situations
– such as at a business social function and at a bar with a group
of friends.
The conjoint attributes and levels are identical, but respondents' resulting
profile preference ratings may be different based on the situation.
If you derive importances for each attribute for each of those two occasions
for the respondents, you will get two sets of derived importances for
each respondent.
There is no reason you cannot subject both sets of derived importances
for these respondents to a standard clustering routine. The same respondent
may then show up in two different clusters, depending on the results
from his situational preferences.
Occasion-Based Segmentation
A particular challenge in market segmentation analysis is how to address
segments when circumstances or occasions drive product preference and
selection.
For example, a researcher is measuring the relative influence of a set
of brands, product attributes, and price variations for carbonated soft
drinks purchased for immediate consumption in a variety of store settings:
grocery, convenience, mass merchandise, deli, and drug store.
Respondents do a point allocation of importance of each attribute, plus
price and brand name, on influencing their selection for each store
setting that they have experienced in the past ten days. In addition,
respondents provide a demographic and consumption volume profile.
The researcher could execute a clustering of the point allocation data
for each type of shopping trip, thus deriving segments on the basis
of the importance drivers for each store type, separately.
Alternatively, the researcher could submit all of the point allocation
data to a clustering algorithm and find clusters or segments in which
the importance drivers are similar within each cluster and different
between clusters, regardless of the occasion. The resulting clusters
may or may not dif-ferentiate between store types.
Either way, the researcher executed an occasion-based segmentation.
Speculations on the Future of Market
Segmentation Research
It seems that the future for market segmentation research is rather
rosy from the demand side. Market segmentation has taken on an increasingly
important role in business strategy development. Thus, senior management
is demanding more segmentation research as a critical input to the strategic
planning process.
Our ability to accumulate and manage massive amounts of data on customers
and potential customers, aligned with the availability of many more
targeted communications capabilities, ensure that there will be increasing
demand for much more and much finer identifi-cation of target markets
in most product and service categories. The massive use of the Internet
only opens up greater possibilities for target marketing.
There are a few downsides. The need for isolating and defining ever-smaller
target markets will require ever-larger research sample sizes and a
commensurate increase in the costs. Samples must be pristine and projectable
to the larger population. For a while, this will preclude using the
Internet as a respondent recruiting method for segmentation research.
This same demand for finer targeting will force more researchers into
the complexities of multidimensional and occasion-based segmentation
These procedures require more time for analysis and reporting, improved
methods for delivering and managing results, and the need for leveraging
database reporting capabilities.
The ANNs and LCMs will continue to supplant many traditional segmentation
algorithms. These require increased methodological and statistical training
for their effective use.
The anticipated changes indicate that the implementation of a segmentation
strategy will get much more complex for both marketers and researchers.
|