C. H. Lien et al. / Asian Journal of Management and Humanity Sciences, Vol. 1, No. 1, pp. 1-15, 2006
Capturing and Evaluating Segments: Using Self-Organizing Maps and K-Means in Market Segmentation CHE-HUI LIEN1,*, ALEX RAMIREZ2, AND GEORGE H. HAINES2 1
Department of Management, Thompson Rivers University, Canada 2 Eric Sprott School of Business, Carleton University, Canada
ABSTRACT Ma r ke ts e g me nt a t i o ni sav i t a lp a r to fa no r g a n i z a t i o n’ sma r ke t i ngbe c a u s ei tprovides the fundamental framework necessary for effective marketing efforts. In recent years, due to their high performance in engineering, artificial neural networks have also been applied in management research. Self-organizing maps, a technique of unsupervised neural networks, are often used for clustering or dimensional reduction. This study employs a modified two-stage approach (SOMs and K-means) to group customers, compares the performance between the tandem approach and direct K-means clustering, and tests for the existence of clusters and segments. The test results show that a media promotion variable would be a basis for segmentation. Based on the segmenting results, a marketing c o mmu ni c a t i o ns t r a t e g yi spr e s e nt e dt oc o pewi t hc us t o me r s ’e x pe c t a t i o ns. Key words: market segmentation, cluster analysis, data mining, neural networks, self-organizing maps.
1. INTRODUCTION Since the pioneering research of Wendell Smith (1956), the concept of market segmentation has been one of the most pervasive in both the marketing academic l i t e r a t u r ea n dpr a c t i c e( Ku o,Ho,& Hu ,2002) .I nt oda y ’ sc ompe t i t i v ema r k e t pl a c e , locating and effectively targeting unique market segments enables a company to understand the wants and needs of its customers. For several decades, statistical cluster analysis has been successfully used in market segmentation (Green & Krieger, 1995). Recently, due to an increase in computer power and a decrease in computer cost, a great deal of interest and effort have been directed towards using neural networks (NNs) in business practice, which were once reserved for multivariate statistical analysis. In marketing, the major application of NNs is on market segmentation. Along with the evolution of data mining techniques, Self-organizing maps (SOMs) have been used in determining clusters and is an alternative approach to statistical clustering techniques (Bigus, 1996; Fish, Barness, & Aiken, 1995; Kuo et al., 2002; Venugopal & Baets, 1994). An SOM provides a mapping from a high-dimensional input data into the lower dimensional output maps. A distinguished feature of the SOM is that it preserves the topology of the input data from the high-dimensional input space onto the output map in such a way that the relative distances between input data are more or less preserved (Garson, 1998; Thanakorn, 2003). Although a number of clustering methods have been presented to solve the market segmentation problem, the importance of testing the validity of clusters is * Corresponding author. E-mail:
[email protected].
1
C. H. Lien et al. / Asian Journal of Management and Humanity Sciences, Vol. 1, No. 1, pp. 1-15, 2006
frequently ignored by marketing researchers (Arnold, 1979; Engelman & Hartigan, 1969; Pilling, Crosby, & Ellen, 1991; Punj & Steward, 1983). If identified clusters are inadequately validated, such clustering results can be considered random, i.e., devoid of a meaningful structure (Pilling et al., 1991; Punj & Steward, 1983). The neglect of the test for clusters could lead to the inclusion of clusters that really do not exist. Segments are defined as groups showing some patterns of similarity and that differ significantly in their response to the relevant marketing variables (Sommers & Barnes, 2003). Clusters identified that differ in some way, such as attitudes or demographics, but not in behavior are not true segments (Massy & Frank, 1965). Kuo et al. (2002) proposed a modified two-stage approach (SOMs and K-means) and, based on simulated data, they found that it is slightly better (lower rate of misclassification) than the traditional statistical clustering method. But Kuo et al. (2002) used a tandem approach, based on a preliminary factor analysis followed by a clustering of rotated, standardized factor scores, in the K-means procedure. Green and Krieger (1995) criticized the tandem approach and they argued that the performance of the tandem approach was not as good as direct K-means clustering. In addition, Kuo et al. (2002) neither tested for the existence of clusters nor evaluated whether their clusters were segments. More empirical studies are needed to support the better performance of the novel two-stage approach. This study uses practical data from the home heating system market and a ppl i e s Ku oe ta l . ’ st wo-stage method (SOMs and K-means) in market segmentation, compares the performance of direct K-means clustering with the tandem approach, and tests for clusters and segments. The remainder of this paper di s c u s s e s SOMs ’a ppl i c a t i on si n ma r k e t segmentation and the test method of clusters and segments, followed by a review of the methodology used in this study. The section after that presents our results. The final section gives a summary and a discussion about the findings of the study.
2. REVIEW OF RELATED LITERATURE 2.1 SOMs in Market Segmentation The general idea of segmentation is to group items that are similar (homogeneous) (Kuo et al., 2002; Smith, 1956). Traditionally, statistical clustering techniques have been the common tools for market segmentation (Green & Krieger, 1995). Punj and Steward (1983) suggested that the integration of a hierarchical a ppr oa c h ,s u c ha sWa r d’ smi n i mum v a r i a n c e ,a l on gwi t han on -hierarchical one, such as the K-means, can provide a better answer than using either a hierarchical or a non-hierarchical method alone. Their approach is called the two-stage approach. The term neural networks arose from artificial intelligence research, which attempted to understand and model brain behavior (Berry & Linoff, 2000; Berson, Smith, & Thearling, 2000). Self-organizing maps are feed-forward, unsupervised neural networks and were developed by Kohonen (Bigus, 1996; Kohonen, 2001). Feed-forward networks are used in situations where we can bring all of the
2
C. H. Lien et al. / Asian Journal of Management and Humanity Sciences, Vol. 1, No. 1, pp. 1-15, 2006
information to bear on a problem at once, and we can present it to the neural networks (Bigus, 1996). Unsupervised learning means that clustering proceeds without knowing (a priori) the number of clusters of the input data, and the network organizes and learns the input data unsupervised (Bigus, 1996). The SOM consists of two layers of processing units, an input layer fully connected to an output layer. There is no hidden layer. Vanugopal and Baets (1994) argued that unsupervised neural networks, including SOMs and the adaptive resonance theory, could be used in determining clusters. Their network for segmentation was conceptually developed. Fish et al. (1995) also suggested that unsupervised neural networks could be utilized for market segmentation. In the housing market, Kauko, Hoomineijer, & Hakfoort (2002) examined the application of neural networks to housing market segmentation in Helsinki, Finland. Their study shows how it is possible to identify various dimensions of housing segments by uncovering patterns in the data set as well as the classification ability of SOM-LVQ (learning vector quantization). Kuo e ta l( 2002 ) ,ba s e donPu n ja n dSt e wa r d’ s( 1983)r e s e a r c h ,pr opos e damodi f i e d two-stage approach, which first used self-organizing maps to determine the number of clusters and then employed the K-means method to find the final solution. Kuo et al. (2002) used simulated data and found that their proposed two-stage approach outperformed the conventional two-stage method. The main reason was that the first stage of the conventional two-stage clustering method always involved the hierarchical methods. One of their shortcomings was the aspect of non-recovery: once an observation has been assigned to a cluster, it should not be moved at all (Kuo et al., 2002). However, SOMs are a kind of learning algorithm, which can c on t i nu a l l yu pda t e ,orr e a s s i gnt h eobs e r v a t i ont ot h ec l os e s tc l us t e r .I nKu oe ta l . ’ s study, they employed the tandem approach in the K-means procedure and the tandem approach was criticized in that it could distort the original cluster structure (Green & Krieger, 1995). It is necessary to examine the non-tandem approach, such as direct K-means clustering, and to evaluate its performance. 2.2 Test of Clusters and Segments An appropriate test for clusters appears to be one, which takes into account the objective of cluster analysis, i.e., to minimize the within-group variance and maximize the between-group variance (Arnold, 1979). Among the test methods, a method suggested by Arnold (1979) appears to be better than other methods ( Pi l l i n ge ta l . ,1991;Pu n j&St e wa r d,1983) .I nAr n ol d’ st e s t ,t h es i gn i f i c a n c eoft h e cluster solution is tested by comparing the calculated value of the C test statistics3 to the values which would be expected if the data were drawn from either a unimodal or a uniform distribution, i.e., no basis for clusters (Arnold, 1979). There is some consensus in marketing research on what makes a segmentation solution a good one (Bacon, 2002; Wedel & Kamakura, 2000). The important criteria are: identifiability, accessibility, substantiality, stability, responsiveness,
3
C = log (max (|T| / |W|)) where T represents the total scatter matrix, and W the pooled within-group matrix.
3
C. H. Lien et al. / Asian Journal of Management and Humanity Sciences, Vol. 1, No. 1, pp. 1-15, 2006
and actionability (Bacon, 2002; Wedel & Kamakura, 2000). These criteria are in terms of usefulness for managerial decision-making and are not easy to evaluate (Bacon, 2002). They are also independent of the technology used to obtain them (Bacon, 2002). Therefore, they provide a useful set of goals for segmentation solutions developed using neural network techniques (Bacon, 2002). Few studies attempted to assess segmentation solutions with these criteria (Wedel & Kamakura, 2000). In this paper, we employ one criterion –responsiveness (consumers in different segments should behave differently toward marketing programs directed at them) –to evaluate whether clusters are segments. The major reason for choosing responsiveness as the criterion is that the data we used are cross-sectional. Other criteria, e.g., stability, need longitudinal data. Responsiveness is consistent wi t hMa s s ya n dFr a nk’ s( 1965)a r g ument, i.e., different segments should have different promotional or marketing variables elasticities.
3. METHODOLOGY 3.1 Research Sample This study used previously collected data (Ratchford & Haines, 1986) for clustering analysis. Data were collected from a US sample of Market Facts Consumer Mail Panel in January-February 1983. Because weather is freezing in winter, home heating systems are important in some states (e.g., New York State) in the US. In the 1980s, basically there are five types of home heating system: gas, oil, electric, heat pump and solar. The families in the US can choose one of them to pr odu c eh e a ti nt h eh ous e .I n Ra t c h f or d & Ha i n e s ’ sr e s e a r c h ,t h e yi de n t i f i e d fourteen perceptual attributes as the criteria for rating the five types of home heating systems. They also run factor analysis among the fourteen attributes and found that the resulting three factors (cleanliness, reliability, and cost) explaining approximately 60% of the variation between attributes. However, they did not perform segmentation among the consumers. We focus on consumers planning on buying a house within the next five years. This research defines the respondents without planning to replace home heating systems in the next five years, but planning on buying a house within the next five years as the potential market. When buying a new house, one type of home heating system must be considered in the purchase. The subjects included both males and females, above the age of 18. 433 responses were obtained for the potential market. 3.2 Research Hypotheses 3.2.1 Test of clusters According to the procedure for testing for clusters described by Arnold (1979), there are two sets in the procedure. The first tests the null hypothesis that the population entities tend to concentrate at one point (i.e., one normally distributed) against the alternative hypothesis that its entities are either uniformly distributed or grouped into clusters. If the null hypothesis is rejected at some level of confidence,
4
C. H. Lien et al. / Asian Journal of Management and Humanity Sciences, Vol. 1, No. 1, pp. 1-15, 2006
a second test is made. The second tests the null hypothesis that the population entities are uniformly distributed against the alternative hypothesis that they form two or more clusters. The two null hypotheses are: H1: the population entities tend to concentrate at one point; H2: the population entities are uniformly distributed. Rejection of H1 and H2 indicates that there are clusters in the population. 3.2.2 Test of segments The null hypothesis is: H3: the marketing variables elasticities among clusters are the same. Rejection of H3 means the clusters are segments. 3.3 Questionnaire Young, Ott, and Feigin (1978) argued that segmentation based on benefits desired is usually the most meaningful type to use from a marketing standpoint as it directly facilitates product planning, positioning, and advertising communication. The main strength of benefit segmentation is that the benefits sought by customers will lead to a causal relationship to future purchase behaviour (Minhas & Jacobs, 1996). This study uses benefit variables to group customers. In Ratchford and Ha i n e s ’ sr e s e a r c h(1986), the questionnaire solicited information on five types of home heating systems and each system type was rated for each of the fourteen attributes (benefit variables, see Table 1). Table 1. Fourteen benefit variables Fourteen Benefit Variables Reliability against breakdown Future availability of fuel supply Floor space required for the system Efficiency of converting fuel source into heat Cleanliness of operating the system Ease of conversion to another system Absence of fumes and odors Absence of pollution Availability of professional servicing Noise-free operation of the system Safety of the system Initial purchase price Warranty protection Annual operating cost
Purchase intentions were measured by the fol l owi ngqu e s t i on :“ I fy ouwe r e buy i ngah e a t i n gs y s t e mf ory ou rh ome ,pl e a s e‘ x’t h eon et y pet h a ty ouwou l dbe mos tl i k e l yt opu r c h a s e . ”Re s pon s e swe r el i mi t e dt ot h ef i v es y s t e mt y pe s .Ot h e r questions involve the knowledge of heating system, rating all the methods of central heating, marking the preference for one method of home heating over the other, media, demographic information, etc. The measurement scale in the questionnaire is 7-point scale, ranged from not very important (=1) to very important (=7).
5
C. H. Lien et al. / Asian Journal of Management and Humanity Sciences, Vol. 1, No. 1, pp. 1-15, 2006
3.4 Generating Clusters: Using SOMs The Kohonen algorithm can be summarized in the following steps (Bigus, 1996; Garson, 1998; Kohonen, 2001) and the self-organizing maps parameters used in our study are listed in Table 2. (1) Neuron weights are initialized to random values. (2) Data representation. When clustering data with neural networks, it is standard practice to scale (or normalize) the input data to a range of zero to one (Bigus, 1996; Garson, 1998). This study uses the sigmoid function4 to transfer input data to a range of zero to one. (3) Setting the learning rate. The learning rate is used to control the step size in the adjustment of the connection weights. This study sets the learning rate as 0.1. The learning rate will decrease over time. (4) Setting the neighborhood. We set a 10-by-10 output layer. The neighborhood is 10 in this study. (5) Presetting the number of training epochs (training cycle). An epoch is the number of training data sets presented to the model between updates of neural weights (Garson, 1998). The purpose of training is that neural networks have the ability to learn and recognize any important pattern or relationship for a set of data (Nguyen & Ramirez, 1998). So far, there is no rule for determining the number of learning epochs. The only way is trial and error (Kuo et al., 2002). In our research, the training epochs for potential market are 80. At the 80th epoch, the network converges to an average total minimum distance. (6) Data points are input to the net, selected at random. (7) Determine which neuron is the least distant from the presented observations. This is the neuron whose weight vector is closest to the input vector from the current observations, measured in Euclidean distance. (8) Weights of neurons in the neighborhood of the winning neuron are adjusted in value to become closer to the value of the winner. The neighbourhood starts out widely defined but decreases spatially as learning iterations proceed, eventually reaching zero (that is, only the weights of the winning neuron are adjusted). When the neighbourhood drops to one, the convergence phase begins. In reviewing the SOM literature (Berry & Linoff, 2000; Bigus, 1996; Kuo et al., 2002), usually 70% of the population used for training, and 30% of the population used for testing were suggested. In the 433 observations of potential market, a sample size of 300 is used for training and the remaining 133 observations are used for testing (see Table 2).
4
The sigmoid function is defined as a strictly increasing function that exhibits smoothness and asymptotic properties. One popular sigmoid function is: f(x) = 1/[1+ exp (-x)].
6
C. H. Lien et al. / Asian Journal of Management and Humanity Sciences, Vol. 1, No. 1, pp. 1-15, 2006
Table 2. The self-organizing maps parameters The self-organizing maps parameters [ 0] Number of Input Unit (
