City University of Hong Kong

CityU Institutional Repository >
3_CityU Electronic Theses and Dissertations >
ETD - Dept. of Management Sciences  >
MS - Master of Philosophy  >

Please use this identifier to cite or link to this item:

Title: Comparison of segmentation methods in services marketing for small to medium enterprises
Other Titles: Zhong xiao xing fu wu ye shi chang xi fen fang fa de bi jiao
Authors: Au, Sai-ming (區世明)
Department: Dept. of Management Sciences
Degree: Master of Philosophy
Issue Date: 2000
Publisher: Dept. of Management Science[s], City University of Hong Kong
Subjects: Market segmentation
Small business -- Marketing
Notes: 126 leaves : ill. ; 30 cm.
CityU Call Number: HF5415.127.A9 2000
Includes bibliographical references (leaves 122-126)
Thesis (M.Phil.)--City University of Hong Kong, 2000
Type: Thesis
Abstract: Market segmentation is a very important and applicable concept in marketing. The small-to-medium enterprises (SMEs) find segmentation strategy particularly useful, as they are closer to their customers. A lot of tools are available for doing market segmentations. However, most of these tools are not specifically designed with the needs and constraints of the SMEs in mind. In addition, many of them are designed to work with large amount of data and it is doubtful if the SMEs can use them. The SMEs do not have the time and resources to develop the best model for segmentation, and they have to trade off between the information quality and the available resources. Survey data are their most important information source as they are unlikely to have large customer database. Thus the SMEs need to find a 'quick-fix' solution that relies on ad hoc collection of data through survey for developing their market segmentation models. It is helpful if they know which classifiers are the most useful for such purpose. From the literature review, it is apparent that there is no universal agreement about the best classification methods to recommend for all occasion, as whose performance is affected by the data types and the application conditions. As classifiers perform best with different types of data, the real research goal for many comparison studies is not to determine the best method in general. Instead the researchers should find out which method works best for some specified data sets (along with their assumptions about the noise, outliers, etc.). In this study, we focus on comparing the performances of various statistical tools and artificial neural networks (ANNs). It is our chief objective to determine which of them are more useful for solving market segmentation problem for the SMEs with survey data. There are many studies comparing different classifiers using continuous multivariate normal data. They are not so useful for finding out the best classifiers from survey data that comprise of both continuous and discrete variables. For instance, in our real data set using the data from a survey for the package tour industry, we find that the segmentation variables comprise of mixed data type with a multivariate normal, a Poisson and a point binomial distribution. Moreover, most previous data for comparing classifiers do not contain noises or outliers. Such comparisons therefore cannot be applied to real-life situations that usually deal with data contaminated with noise and outliers. Noises are common as they represent measurement errors (e.g. the respondents have recalled the frequency of purchase wrongly). The outliers arise occasionally and they are simply observations that do not belong to any classes. To get a good segmentation result we need a classifier that is robust to both noises and outliers. To determine the best classifiers, it is useful to compare the classifiers with both real data and simulated data. We use real-life survey data to provide information for the overall data structure and contextual meaning for the identified segments that is useful for the evaluation of segments in a practical sense. Real-life data are useful in showing how effective these segmentation tools are in providing the useful information for the development of market segments. On the other hand, data from the Monte Carlo simulation are used to compare the efficiency of true class recovery by different classifiers. Simulated data with parameters determined from real-life data are useful for choosing the best classifiers as their true classes are known beforehand. In our study, comparisons were made using simulated data with carefully selected parameters for different experimental designs. Statistical significance can be attached to the tests and the results thus can be generalized to similar situations. Based on simulated data sets, we performed a comparative study of eight classifiers, namely the unsupervised learning classifiers including the K-means, Autoclass, finite mixture model and the self-organizing map; the supervised learning classifiers including the linear discriminant analysis, decision tree, backward propagation neural network and the basis radial function neural network. To mimic the real situations faced by the SMEs, the experimental design involves different levels of factors including number of classes, class probability distribution, noise levels and outliers levels and the data sets are generated accordingly. Managers may choose unsupervised or supervised learning for market segmentation. Unsupervised learning is useful for discovering patterns or clusters amongst the data. We find that amongst the unsupervised learning classifiers, the Autoclass and mixture model perform most satisfactorily when using accuracy and Adjusted Rand as the performance measurement. We recommend using these two methods in unsupervised learning to cluster the data for segment identification; each serves as the external criterion for validating the segments formed by the other method. The determination of number of classes is a difficult problem and Autoclass is useful for determining the actual number of classes in noise-contained situations. Supervised learning can be used for predicting class membership, for instance in the application in direct marketing. It can also be used to fine-tune the result from unsupervised learning or to improve the interpretability of the resulting segments. It is found that discriminant analysis and classification tree perform better than the other supervised learning classifiers. Besides, the classification tree is credited for providing visual presentation to show the problem of complexity relating the independent variables to the dependent variable. In a nutshell, we recommend SMEs to use Autoclass and mixture model for unsupervised learning and discriminant analysis and classification tree for supervised learning for market segmentation using survey data. However, we believe that the SMEs, with the help of some market research professionals, can be further benefited from our study through our proposed process model for market segmentation. Our process model is more general and is needed because the performance of the classifiers is problem specific. The segmentation result may not be optimal if some of our assumptions about the data are not true, e.g. the researchers may have included irrelevant variables for segmentation. Our process model tells them how to perform segmentation from scratch. Therefore we extend our findings by proposing a market segmentation process model for the SMEs. Our process model helps the researcher determine the best segmentation starting from survey data. At first, we use the Structural Equation Model (SEM) to decide on what variables are useful for segmentation analysis. The inclusion of the relevant variables for segmentation is important since the inclusion of irrelevant variables will adversely affect the segmentation result. The SEM enables the researchers to limit the variables included in the segmentation using causal paths. Then both unsupervised learning and supervised learning classifiers should be used for market segmentation. The former looks for patterns in the data to form segments exploratorily and the latter tells us how the segments differ from each other and can refine the segments. To allow for the possibility of having different classifiers to suit different data structures, simulated data with known parameters determined from the real data set are used to find out which classifiers give the best results. Finally, the best classifiers are used to segment the real data. Depending on the level of competence of the users, they may start from scratch and follow through all stages in our process model. Alternatively, if they have mixed data with multivariate normal, Poisson and point binomial distributions, they may apply our results directly. They can first use Autoclass and Mixture model for the unsupervised learning. The model can then be fine-tuned or better understood by using the classification tree and discriminant analysis. In either way, they will find our results very useful for performing market segmentation.
Online Catalog Link:
Appears in Collections:MS - Master of Philosophy

Files in This Item:

File Description SizeFormat
fulltext.html159 BHTMLView/Open
abstract.html159 BHTMLView/Open

Items in CityU IR are protected by copyright, with all rights reserved, unless otherwise indicated.


Valid XHTML 1.0!
DSpace Software © 2013 CityU Library - Send feedback to Library Systems
Privacy Policy · Copyright · Disclaimer