An extended self-organizing map network for market segmentation—a telecommunication example

https://doi.org/10.1016/j.dss.2004.09.012Get rights and content

Abstract

Kohonen's self-organizing map (SOM) network is an unsupervised learning neural network that maps an n-dimensional input data to a lower dimensional output map while maintaining the original topological relations. The extended SOM network further groups the nodes on the output map into a user specified number of clusters. In this research effort, we applied this extended version of SOM networks to a consumer data set from American Telephone and Telegraph Company (AT&T). Results using the AT&T data indicate that the extended SOM network performs better than the two-step procedure that combines factor analysis and K-means cluster analysis in uncovering market segments.

Introduction

Market segmentation refers to the process of forming groups of consumers, whereby the groups are homogeneous in terms of demand elasticities and are accessible via marketing strategies [13]. A marketing manager can select and effectively execute segment-specific marketing mixes. The value of performing market segmentation analysis includes positioning the product in the marketplace properly, identifying the appropriate segments for target marketing, finding opportunities in existing markets, and gaining competitive advantage through product differentiation. The bottom line is to increase profitability by enabling firms to target consumers more effectively. Although it was introduced into the academic marketing literature in the 1950s, market segmentation continues to be an important focal point of ongoing research and marketing practices [4]. Most of the academic research in market segmentation has been in the development of new techniques and methodologies for segmenting markets. The common thread running through these diverse streams of research is the attempt to segment consumers, deterministically or probabilistically, into a finite number of segments that are homogenous within and heterogeneous between with respect to consumer demand and demographics.

It should be noted that the usefulness of market segmentation hinges upon the accuracy of market segmentation. Relatively low accuracy in forecasting segment memberships indicates a high portion of unintended members in each segment. Misplacements will result in ineffective marketing programs that are designed to stimulate sales as well as in potential negative impact on revenue generation from the unintended segment members.

From a methodological point of view, a clustering procedure is often employed to form segments of consumers with similar preferences. When the number of dimensions underlying preferences is large, a researcher may first use a dimension reduction technique such as principal component or factor analysis to reduce the dimensions to a manageable set before subjecting the output factors to a clustering routine. Thus, a two-step approach is typically used [29].

Statistical dimension reduction routines such as factor analysis evolve around the Pearson's product moment correlation coefficient. Multivariate normality and linearity among the variables are key assumptions underlying the use of correlation coefficient. Clustering algorithms are in general heuristics in that they assign objects into clusters based on some distance measures between an object and the centroid of the cluster. Clustering algorithms are not statistical in the sense that they do not rely on any distributional assumptions. Violation of statistical assumptions in this two-step procedure will likely come about in the data reduction stage. In commercial market research applications, measures such as consumer product ratings, customer satisfaction assessments tend to be markedly skewed [26]. Violations of the normality assumption may lead to bias and incorrect assignment of consumers to the resulting segments and eventually lead to ineffective marketing strategies.

Kohonen's [10], [11], [12] self-organizing map (SOM) network, a variation of neural computing networks, is a nonparametric approach that makes no assumptions about the underlying population distribution and is independent of prior information. Similar to principal component analysis and factor analysis, the main function of an SOM network is dimension reduction; that is, it maps an n-dimensional input space to a lower dimensional (usually one- or two-dimensional) output map while maintaining the original topological relations and, thus, enables decision maker to visualize the relationships among inputs. While Kohonen's Self-Organizing networks have been successfully applied as a classification tool to various problem domains, including speech recognition [15], [32], image data compression [17], image or character recognition [2], [25], robot control [23], [28], and medical diagnosis [27], its potential as a robust substitute for clustering tools remains relatively unexplored. Given the nonparametric feature of SOM, it can be expected that SOM will yield superior results to the factor/cluster procedure for market segmentation. Recently, neural network applications for clustering and prediction in marketing [3], [7] showed promising results as compared to the traditional statistical approaches.

Balakrishnan et al. [1] compared several unsupervised neural networks with K-means analysis. Due to the lack of an extended grouping function such as the one implemented in our extended SOM network, the two-layer Kohonen network implemented in their study is designed so that the number of nodes in the output layer (Kohonen layer) corresponds to the number of desired clusters. It is a different kind of Kohonen network that does not provide a two-dimensional map for users to visualize the relationships among data points. We have found that the majority of the studies that apply Kohonen network to clustering have implemented this type of network. The performance of these neural networks is examined with respect to changes in the number of attributes, the number of clusters, and the amount of error in the data. Their results show that the K-means procedure always outperforms neural networks especially when the number of clusters increases from two to five. Unlike the compromising approach of Balakrishnan et al., our extended SOM method preserves the dimension reduction function of the original SOM and further groups the nodes on the output map into a user specified number of clusters. The extended SOM will be discussed and presented in detail in a later section. Key features specific to the expended SOM will be pointed out. How these added features serve to enhance the performance of SOM in dimension reduction and clustering will also be examined.

In this study, we will first employ the extended SOM to group consumers into segments using their attitudes toward long-distance communications. The resulting segments will be cross validated using consumer usage of four primary modes of communications (long-distance phone calls, letters, cards, and personal visits) and demographic factors. The key demographic factors identified by American Telephone and Telegraph Company (AT&T) are number of friends (FRIENDS) and relatives (RELATIVES), number of moves made in the past 5 years (MOVES), number of people over 16 years of age currently living in the household (OVER16), and marital status of the head of household (MSTATUS). The first two factors correspond to the size of community of interest. All of these demographic factors have been found by the company to be strongly correlated with the usage rate of long-distance phone calling. Results from using the extended SOM will be compared with those from the factor/cluster procedure.

The balance of the paper is organized as follows: Section 2 presents the basic concepts of SOM networks and illustrates their use as a dimension-reduction tool (analogous to factor analysis). This is followed by a discussion of the extended grouping capability integrated into the original SOM networks. Section 3 describes the experimental procedures and results from AT&T data set. In Section 4, we compare the performance of the extended SOM with that of the factor score-based approach, both qualitatively and quantitatively. The segments will be validated using consumer usage and demographic factors. The paper concludes with a summary of our findings.

Section snippets

Self-organizing map (SOM) networks

The self-organizing map (SOM) network is a neural network based method for dimension reduction. SOM can learn from complex, multidimensional data and transform them into a map of fewer dimensions, such as a two-dimensional plot. The two-dimensional plot provides an easy-to-use graphical user interface to help the decision-maker visualize the similarities between consumer preference patterns. In the AT&T data set, there are 68 customer attitude variables. It would be difficult to visually

The empirical study

There are three main components identified in the market segmentation literature including: formation of market segments, differences across segments with respect to consumer demand, and identification of the segments using consumer demographics. The factors used in the formation of segments are referred to as the base variables. In our example, these are the 68 attitude items. Criterion variables correspond to the consumer demand/preferences within each segment. Again, these are expected to be

The comparative study

In market segmentation studies, the accurate assignment of respondents to clusters/segments is critical. Kiang and Kumar [9] used simulated data where true cluster memberships are known and found that SOM networks provide more accurate recovery of underlying cluster structures when the input data are skewed. Because true cluster memberships are unknown for the real-world data set, it is difficult to determine which clustering method is best for the real-life problem. The accuracy of membership

Conclusion

In this study, we first identify the importance of market segmentation in market management and strategy development. The first and primary component in market segmentation is the formation of groups. Accurate assignment of segment membership is the key to successful market segmentation programs. Traditional approach relies heavily on factor analysis for dimension reduction and cluster analysis for grouping. The imposition of linearity and normality assumptions may lead to less-desirable

Melody Y. Kiang is a Professor of Computer Information Systems at California State University, Long Beach. She received her MS in MIS from the University of Wisconsin, Madison, and PhD in MSIS from the University of Texas at Austin. Prior to joining CSULB, she was an Associate Professor at the Arizona State University. Her research interests include the development and applications of artificial intelligence techniques to a variety of business problems. Her research has appeared in Information

References (32)

  • M.Y. Hu et al.

    Explaining consumer choice through neural networks: The Stacked general approach

    Journal of Operational Research

    (2003)
  • L. Hubert et al.

    Comparing partitions

    Journal of Classification

    (1985)
  • M.Y. Kiang et al.

    An evaluation of self-organizing map networks as a robust alternative to factor analysis in data mining applications

    Information Systems Research

    (2001 (June))
  • T. Kohonen

    Cybernetic Systems: Recognition, Learning, Self-Organization

  • T. Kohonen

    Self-Organization and Associative Memory

    (1989)
  • T. Kohonen

    Self-Organizing Maps

    (1995)
  • Cited by (0)

    Melody Y. Kiang is a Professor of Computer Information Systems at California State University, Long Beach. She received her MS in MIS from the University of Wisconsin, Madison, and PhD in MSIS from the University of Texas at Austin. Prior to joining CSULB, she was an Associate Professor at the Arizona State University. Her research interests include the development and applications of artificial intelligence techniques to a variety of business problems. Her research has appeared in Information Systems Research (ISR), Management Science, Journal of Management Information Systems, Decision Support Systems, IEEE Transactions on SMC, EJOR, and other professional journals. She is an Associate Editor of Decision Support Systems and Co-Editor of Journal of Electronic Commerce Research.

    Dr. Michael Hu holds the Bridgestone Endowed Chair in International Business and he is a Professor of Marketing at Kent State University. He earned his PhD from the University of Minnesota in Management Science/Marketing. He is a dedicated educator and won the University Distinguished Teaching award in 1994. He has published over a hundred and ten journal articles in the areas of artificial neural networks, international business and marketing research. His research has appeared in Decision Support Systems, Annals of Operations Research, European Journal of Operational Research, Decision Sciences, Journal of Marketing Research, Journal of International Business Studies, among many others.

    Dorothy M. Fisher is a Professor of Information Systems at the California State University, Dominguez Hills. She received an MA from Duke University and a PhD from Kent State University. Dr. Fisher has had broad consulting experience with private firms as well as educational institutions. Her research emphasizes the applications of statistical and artificial intelligence techniques to management problems. Dr. Fisher has published papers in the Journal of Computer Information Systems, the Journal of Systems Management, the Journal of Applied Business Research, and other academic and professional journals. Currently, she is the Managing Editor of the Journal of Electronic Commerce Research.

    View full text