Words: 3457
Presented segmentation approach comprises classic loyaltyprofitablllty link model that Is explicit for CRM, and ewe social media components direct to Social CRM. The result of presented approach is an analysis tool with data visualization for managers which significantly improves the process of customer segmentation, Presented research Is supported by implementation of proposed approach by which experiments were conducted. Additionally, the experimental results showed that proposed method performed very close to k-means algorithm which Indicate the correctness of the proposed approach.
Smokers: customer segmentation, CRM, Social CRM, classification, SOME, unsupervised learning, ANN, data mining. Introduction To acquire competitive advantage many companies use the strategy of Customer Relationship Management (CRM) what can be observed in growing Interest In this domain. However, in recent years new element of strategic importance appeared called social media. In order to meet the changing expectations of customers needs, Social CRM (SCRAM) systems represent new branch of CRM systems which is oriented on the use of social media.
With the emergence of a new family of CRM systems, there has arisen the need for developing tools supporting these systems. Although both CRM and SCRAM systems eve many analytical tools, still a lot of them impose the necessity of extensive data management and using external software packages [1]. This in turn causes that many analyses are carried out semi-automatically or even manually. Such a situation results in not only a loss of valuable time, but also a lack of focus on the most important components of customer relationship management systems and their benefits.
This comes down to the fact that the management Is unable to keep up with the rapidly changing customers trends, particularly in the area of social networks. A. K;octet, P. Q, and P. Steer (Des. ): CNN 2013, case 370, up. 552-561, 2013. C Automatic Customer Segmentation for Social CRM Systems 553 The aim of this work is to propose an approach to solve the problem of automatic segmentation of customers in the SCRAM systems. The purpose of the method is to support the CRM strategy by providing applicable tools of data analysis for managerial staff.
Presented segmentation approach is based on well-known model, linking customer profitability and loyalty, which are also the two most important components of CRM strategy. Moreover, the presented approach has been extended to include elements elated to social media, which are crucial to SCRAM systems. It was also assumed that each of the main segmentation components can be composed of many features. In addition, an adequate representation of analyzed data that provides management with clear results in form of diagrams is required.
Therefore, for customer segmentation the Self Organizing Maps (SOME) algorithm is proposed which is commonly used for classification and visualization of high-dimensional data. 2 Related Work This paper is a continuation to our research on intelligent tools supporting CRM systems using their information potential. The research presented in [2] includes definitions of some indicators, which were also used in this paper. This includes RFM (Regency Frequency Money), LTV (customer LifeTime Value), and NP (Next Purchase Probability) used for customer loyalty estimation.
The theoretical foundation on classification using SOME is [3] whereas the basic concepts for customer segmentation using data mining techniques in CRM are presented in [1]. There are also numerous studies devoted to the area of customer segmentation in CRM. Some of them use genetic algorithms (for example [4]), however this technique is not eatable for automatic segmentation since it requires lot of data-specific parameterization and complex fitness function for high-dimensional data. On the other hand, most of the studies on customer segmentation for CRM focus on solving classification or classification task.
They can be divided into three main classes according to the data mining model they apply: (I) supervised learning models, (ii) unsupervised learning models, (iii) and hybrid models. First group which addresses the classification problem (e. G. [5]) could be applied for segmentation of high-dimensional data or even take into account imbalanced data 6]. However, classification requires establishing a training set that will represent particular segments. For this reason the classification methods are not applicable for automatic segmentation.
Second group that use classification techniques seem to be appropriate for resulting segments are not known in advance. In [7] authors presented intelligent method which uses SOME. However, the whole process of segmentation is preceded by linear programming method, of which output is later used for SOME training. In such a case SOME is used as a classifier. In result the method requires setting a few ramset’s for segmentation (for linear programming and 554 A. Cozy’s and A. Crazy SOME classifier) and it cannot be applied for automatic segmentation.
Moreover, the method puts focus on profitability only, which seems not to treat the problem of segmentation in CRM as a whole. Another study on segmentation using classification techniques was presented in [8] and [9], however both methods require that some of the parameters should be determined empirically, which limits their automation capabilities. The last group represents methods that combine supervised learning models with unsupervised ones. In [10] authors presented segmentation method utilizing both SOME and back-propagation ANN (Artificial Neural Networks).
The approach itself proposes a classifier where part of the input data comes from SOME. Although this method is not suitable for automatic segmentation, it allows expanding the number of features. However, the feature expansion is possibly at the cost of re- parameterization of the network. All of the above studies use different segmentation methods and different data sets suitable for different purposes. Moreover, vast majority of mentioned studies were application-specific whereas different areas of application affect selection of different segmentation components.
Despite the fact that most of mentioned studies use similar features where RFM and LTV are the most common ones, some of them focus on profitability only while other take it into account only indirectly. In result, it is difficult to compare presented approaches of customer segmentation to each other. Additionally, none of above research considers the Social CRM segmentation problem and none of above works propose data visualization tools as the outcome of their methods. 3 Segmentation Problem segmentation of customers in the SCRAM systems and visualize it in such a manner that it provide a complete tool of data analysis for management.
In order to create automated tools for customer segmentation, presented approach requires intelligent techniques of data mining. One of such a techniques are Self Organizing Maps (SOME) which is commonly used also referred to as Chosen network, is an Artificial Neural Network approach which uses unsupervised learning algorithm. It is composed of a map or grid of neural cells where each of them is associated with a n-dimensional vector. Cells are adjusted through a learning process using n-dimensional training data.
No supervision means that there is no human expert who must assign input data to particular class and provide it as training data. In clustering, it is the distribution of the data that will determine cluster membership. The usage of SOME also gives the possibility to make low-dimensional representation of possibly high-dimensional data set which is crucial for proposed approach of customer segmentation. This is possible because SOME neighborhood function preserves the topological properties of the input data space. Based on the 555 above-mentioned reasons, this method is selected for the approach presented in this paper.
Proposed in this section segmentation methods is based on classic model of customer loyalty-profitability segments. This model became very popular after article published in 2002 by W. Reinsert and V. Kumar [11]. Authors indicated that there is evident link between loyalty and profitability that can by illustrated by Fig. 1 [12]. Fig. 1 . Customer segments in the context of loyalty-profitability link This model assumes that customers are sorted according to their profitability and ingenuity [1 1] where longevity translates to loyalty.
By dividing those two indicators by high and low groups, the resulting four categories represent customer segments to which different marketing strategies should be applied. Since loyalty and profitability are the most important components of CRM strategy, in CRM systems there are measures that allow the calculation of both of those factors. Presented approach uses classic methods to measure the level of customer loyalty which was presented in [2]. This includes calculating loyalty using three features: the RFM Probability).
Moreover, the presented in this paper approach considers modifying and extending the classic segmentation model in order to include elements related Social CRM systems. The modification is based on proposing new method of measuring customer’s profitability in the sphere of social media and it is called Social Media Profitability (SMS). The extension is based on measuring the level of client’s Social Media Engagement (SEEM), which in result adds a new dimension to standard segmentation model presented on Fig. . For the purpose of this paper the calculation of Social Media Profitability is done sing only one feature called Social Media Return of Investment (OSMOSIS). Measuring it into social media seem to be an accurate measure of profitability for SCRAM. The OSMOSIS is calculated per customer as profit from social networking events (for example on Backbone) divided by money invested on those events. The SMS can be later extended with more features to better reflect profitability, especially in the long- term view. 56 The calculation of Social Media Engagement so far also includes only one feature called Weighted Event Participation (WEEP) calculated as customer’s average event articulation multiplied by 1 + Twitter Follow Frequency (OTF). The OTF is used to measure how often a person attracts new followers on Twitter and is calculated as average number of new followers per day. Applied as a multiplier allows to emphasize customer’s influence on others in social networking services. Considering the above, presented segmentation problem can be reduced to hard classification problem which is solved using SOME function.
This can be formally defined as follows: Definition 1. Given a customer data set C = {CLC , co , . , CNN }, where each CIA C is represented by following duple CIA = XSL ,XX,XX . Elements X] . . 3) represent SCRAM segmentation components such as loyalty, SMS, SEEM, and are defined as feature groups X] = {XSL ,xx, . , XML } composed of real-valued XSL e X] feature vector in Euclidean space. Then the clustering of C is the partitioning it into k sets K = {KIWI , . , K } called clusters using an assignments : C -?+ K where s is surjection SOME function (none of the clusters in K is empty).
It is also required that function s is not objective (there is no one to one mapping) SKI and ink=l ˜. Problem definition formulated in such a manner allows to extend feature groups X] to NY size, and thereby increasing the number of dimensions of the input vectors. Therefore, the loyalty, SMS and SEEM measures can be easily expanded by adding more factors and better reflecting models they represent. In addition, the definition of customer data CIA as three-tepees gives the possibility of representation multidimensional data by the means of colors, using the ORG color model.
The second advantage is the ease of handling this type of data on two- and three- dimensional graphs for data visualization used by managers. To do that it is enough to calculate the mean of feature vector components in particular feature group X] , ND scale it into range between O and 255 using min-Max normalization. Clustered data is later plotted on graphs using cluster color to which it was assigned. Three types of graphs are proposed – first one representing data in loyalty relation, second one in loyalty-SEEM relation, and the last one in the loyalty-SEEM relation.
Furthermore, data on first two graphs is divided into halves using vertical and horizontal lines. In result, clustered data is additionally divided into four areas A, B, C, D. This procedure allows to reflect the classical segmentation methodology marketing strategies. Moreover, this allows to compare clusters in terms relation to particular areas. Despite the fact that data visualization using graphs is based on two or three dimensions only, the SOME classification operates on multi-dimensional data.
In this paper the number of dimensions is limited to five, according to presented loyalty, SMS and SEEM schemes. 557 In order to make the segmentation process on the basis of the above model ran automatically, SOME parameter values must be fixed or depend on the parameters of the model. The bigger the map the better the clustering results, however the number f resulting clusters grows and the algorithm execution time significantly increases. Smaller maps are more generalized but number of classes is smaller and execution time is much shorter.
For the purpose of this paper it is assumed that the SOME parameters are fixed and in order to get appropriate outcome for presented segmentation method, that is to keep number of resulting classes reasonably low with good classification quality and low algorithm execution time, the SOME map was proposed to be of size xx, number of interactions equal to 1000 and learning rate set to 0. 5. However, it should be noted that the presented model assumes that the number of features can be arbitrarily increased, which can affect the quality of clustering for SOME with fixed parameters.
In addition, the number of input data may also have impact on the results. Initial empirical studies have shown that for the assumed constant parameter values of SOME the size and number of dimensions of input data had negligible impact on the number of resulting clusters. Such an assumption, however, requires verification and thus assumed SOME constant values are preliminary. Their optimization, finding limit values and determining dependencies to the model are the subject of further research. 4 Experimental Results Based on implementation of presented approach an experiment was conducted.
The aim of the experiment was to apply presented approach of automatic customer segmentation on a case study example. The input data for the experiment consisted of 50 customers collected from SCRAM system of a textile industry company. Clustering results are also evaluated by the means of comparison to k-means algorithm. The data set did not contain any personal information about customers. All collected data concerned one year period and all feature values are calculated according to this timestamp. Given input data was divided into 7 clusters and presented on Fig. 2.
First cluster 3 customers, third (POS. [9, 3]) 2 customers, fourth (POS. [3, 3]) 14 customers, fifth (POS. [O, 13]) 6 customers, sixth (POS. [13, 9]) 2 customers, and seventh (POS. [13, 3]) contained 1 customer. It is important to notice that some clusters that are located near one another and seem to share similar color (as for example clusters with centered POS. [O, 13] and [8, 1 3]) are significantly different because their Euclidean distance is big. Such a situation may happen because SOME map visualization is done using only three parameters (ORG colors), whereas nodes are composed of five features.
The true border between clusters is visible on the I-I-matrix (Unified Distance Matrix) illustrated on Fig. 3 which gave a two-dimensional representation of high- dimensional data (in presented case five-dimensional). The u-matrix presents the distance between the adjacent neurons and depicts it in gray-scale. 558 Fig. 2. Two-dimensional SOME node layer of size xx. Left side of the figure (a) presents map without plotted data, whereas map on the right (b) includes clustered ATA. The more red the more objects particular class contains.
Yellow field indicates that there is only one object in that cluster. A dark color between the neurons corresponds to a large distance and thus is considered as cluster separator. A light color between the neurons indicates that they are close to each other and thus light areas represent clusters. The value of a particular node is the average distance between the node and its closest neighborhood. There are two types of u-matrix graph presented on Fig. 3. The square grid u-matrix allows to present up to 8 neighbors of a node, whereas hexagonal grid allows to consider up to 6 neighbors.
Fig. 3. I-I-matrix SOME representation illustrating distances between neurons (marked with black dots). Left side of the figure (a) presents square grid u-matrix, the middle (b) presents square grid u-matrix with highlighted colors of corresponding clusters, and the right side (c) presents hexagonal grid u-matrix. Subsequent aspects of customers segmentation – the loyalty-SMS link (Fig. AAA), loyalty-SEEM link (Fig. B), and three-dimensional graph of loyalty-SMS-SEEM (Fig. 5). Each segment is automatically assigned a color that corresponds to the color of its cluster.
As it was mentioned this data is expressed in ORG color model, where the first component represents loyalty, the second customer 559 profitability in the area of social media and the third of his/her engagement. As a result, clusters can be represented in the full range of colors. Figure 4 present the customer segmentation results of 50 customers. Fig. 4. Clustered data plotted on loyalty-SMS graph (a) and loyalty-SEEM graph (b). Vertical and horizontal lines divide data by halves. Vertical and horizontal lines divide the data in halves to reflect the classical idea of the segmentation shown in Fig. ND allow to combine it with the identified clusters. As a result, customers are divided into seven segments in eight areas (four areas A, B, C, D for each dimension). A complete picture of resulting segmentation is presented on the graph as shown in Fig. 5. Based on this kind of data visualization, managers can immediately plan separate marketing activities for each segment. For example, customers of the segment [O, 7] in the area C are characterized by relatively low loyalty, low engagement to social media events, and low profitability.
This group can be certainly classified as strangers. For this group, a good strategy might be to encourage more frequent purchases by offering a loyalty program. Much better group of customers belong to segment [8, 13] in the same area. However, their relatively high SEEM is not reflected in profitability. Perhaps these are the customers who prefer on-line shopping. Similar element of classical approach to segmentation presented in [1 1] can be applied to other identified segments.
As can be seen automatic segmentation using SOME clustering technique gives much more detailed client groups than in the case of manual segmentation by determining angles for each segment, especially in the case of multi-dimensional data. Proposed approach allows to save time and better meet customer needs so the company can achieve greater benefits. In the last stage of the experiment the SOME clustering was compared to clustering generated using k-means algorithm with the same number of iterations (1000) and number of clusters (k = 7). The results are presented in Table 1.
The Silhouette Score appropriately the data has been clustered [13]. It Fig. 5. Clustered data plotted in three-dimensional loyalty-SMS-SEEM space Table 1 . Comparison of clustering using SOME and k-means SOME Silhouette Score: K-means Silhouette Score: Normalized Mutual Information: Rand Index Score: Adjusted Mutual Information: 0. 222 0. 238 0. 667 0. 458 0. 538 ranges from -1 to 1 therefore result close to 0. 2 is satisfying. Remaining metrics allow to measure a similarity between two clustering [13]. Comparison results show that SOME classification performed very close to k-means and resulting clustering are relatively similar.
This indicates that proposed approach of customer segmentation for CRM is valid. 5 Conclusions and Future Work The aim of this work was to propose an approach to solve the problem of automatic segmentation of customers in the Social CRM systems using Self Organizing Maps. The purpose of the method was to support the CRM strategy by providing tools of data analysis for managerial staff. Presented segmentation approach was based on well-known model, linking customer profitability and loyalty, which represent the two most important components of CRM strategy.
Moreover, the presented approach included new elements related to social media, which are crucial to SCRAM systems. Presented segmentation model also assumed that each of the main segmentation components can be composed of many features. Additionally, the model allows further feature expansion by adding more factors to loyalty, social media profitability (SMS) and social media engagement (SEEM) components to better reflect information they carry. Based on presented approach of automatic customer segmentation an experiment was conducted in a form of a case study example.
The input data for the experiment considered 50 customers collected from 561 SCRAM system of a textile industry company. Additionally, the experimental results showed that proposed method performed very close to k-means algorithm which indicate the correctness of the proposed approach. However, the advantage of SOME algorithm is that it does not require defining number of clusters which in case of automatic approach is not known a prior’. On the other hand, presented approach requires SOME parameters optimization, finding limit values and determining dependencies to presented classification model.