Travel Reference
In-Depth Information
applying Biclustering. This procedure providing algorithms for all scale levels simultaneously
classifi es variables and cases. For each segment only relevant subgroups of variables are used
while invalid ones are ignored. Analysts may choose whether they want to generate overlap-
ping or non-overlapping clusters, and they can trust that repeated computation runs lead
to identical results. Dolnicar
et al
. (2012) provide empirical evidence for the superiority
of Biclustering by comparing leisure activity segmentation results with results using k-means
and Ward's method.
Since k-means may not result in a global optimum several extensions such as fuzzy k-means
or overlapping k-centroids clustering (Chaturvedi
et al
. 1997) were introduced but have not yet
become popular in tourism marketing. Other extensions of partitioning approaches have found
their way into tourism research. Examples are Bagged Clustering (Dolnicar and Leisch 2000;
Leisch 1999), Vector Quantization techniques and other fuzzy or probabilistic algorithms (e.g.
Rough Set Theory Fuzzy Set Theory or Latent Class Analysis).
Vector Quantization procedures such as the Topology Representing Network (TRN) and
Kohonen's Self Organizing Map (SOM) produce more stable results than ordinary k-means
clustering. According to Martinez
et al
. (1993) TRN is faster and achieves smaller distortion error
than both SOM and k-means. SOM is for instance used by Rong
et al
. (2012) to identify
segments among online users of hotel websites. TRN is a nonparametric method applying the
'neural gas' algorithm (Martinetz and Schulten 1994). It is particularly useful for condensing
three-way data (respondents, objects and their attributes); offers heuristics for interpretive
assistance (Mazanec 2001); and provides methods for estimating the stability of results (see e.g.
Ganglmair and Wooliscroft 2000, who report on a comparison of k-means and TRN proving
that k-means is very sensitive to changes in the order of input data). Lately, TRN got applied to
cluster ranking data concerning quality of life domains (Dolnicar
et al
. 2012).
Tourism research has begun experimenting with probabilistic classifi cation approaches
such as Latent Class Analysis or methods based on Rough Set Theory. Rough set classifi cation
handles non-numeric classifi cation replacing incomplete or imprecise data with precise lower
and upper approximations (Pawlak 1984). Rough set classifi cation was applied in tourism to
induce decision rules which were then used to forecast dining behaviour (Au and Law 2002).
Voges (2006) introduced an evolutionary algorithm based on rough clustering which overcomes
the dependence of k-means on seed points. The algorithm allows objects to be a member
of several clusters. Another approach capable of dealing with vague data is Fuzzy Set Theory
(Zadeh 1965). A fuzzy set implies a function of membership which allocates to each object a
membership grade between 0 and 1. Hsu and Lin (2006) apply this technique for classifying
travellers' risk perceptions.
Validation and replication
External validity may be tested by randomly splitting one's sample and using the fi rst half for
the segmentation and the second half for testing the results (e.g. Tkaczynski and Prebensen
2012). Several clustering methods may be compared, or differences in cluster characteris-
tics/profi les can be examined statistically (e.g. Spotts and Mahoney 1993). Reliability can be
tested by repeating a clustering procedure and judging the stability of results (Dolnicar
2002). The development of segments may also be monitored over time by investigating a
priori segments or by tracking changes of a posteriori segments over time emerging
from repetitive surveys (Dolnicar 2004b). Reference examples for repetitive surveys are rare;
however, Shoemaker (2000) shows how a mature market has changed in the course of a ten-year
time period.
