Data mining methods give opportunity to solve problems of current interest, market segmentation problem belongs to which. T here are different approaches to solve market segmentation problem that differ by used methods. Data mining methods are used to solve classification and clustering problems. T here are k-means method, EM algorithm and neural networks which are considered and compared. Deductor platform is used to analyse implementation of clustering algorithms. Ke ywords: market segmentation, customer preferences, clustering, similarity measure, neural networks.
Development of methods for recording and storing data has currently led to a rapid
growth of the amount of collected information. The volumes of data are so impressive
that it is simply not realistic for a person to cope with data analysis on their own and
the problemto analyze collected information more efficiently is continuously
increasing. All modern companies today have the ability to store data in a single database,
called the customer base [
With a large number of customers, it is difficult to build an individual approach, so it is convenient to group theminto groups with homogeneous characteristics, which are segments. Clustering can be used to segment and build customer profiles. The efficiency of working with clients increases by taking into account their personal preferences. Clustering can be used in a wide variety of areas; they are retail, banking, telecommunications, insurance, government services and others. 2
Clustering differs from classification in that an output variable is not required and the
number of clusters into which a set of data must be grouped may not be known. The
output of clustering is not a ready-made answer. A cluster is a group of similar objects.
Clustering indicates the similarity of objects which it includes. The resulting clusters
require additional interpretation. To determine the similarity of objects, it is needed to
set a measure of proximity. The most popular measure of proximity in two dimensions
is the Cartesian distance or Minkowski metric [
The k-means algorithmis one of the simplest, but at the same time, not entirely a
ccurate clustering method [
In (1) k is the number of clusters, Si are the resulting clusters, i varies from 1 to k and μi are centers of elements x fromcluster Si. The algorithmis performed iteratively until the boundaries of the clusters and the location of the centroids change. The algorithm may take a dozen iterations to execute. The advantage of the algorithm is simplicity of implementation and speed of execution. The disadvantage is the need to initially set the number of clusters and select the initial mass centers.
Kohonen neural networks form a class of neural networks, the main element of
which is the Kohonen layer, they are used for data analysis and solving clustering
problems [
The EM (Expectation Maximization) algorithmis based, as it were, on maximizing
expectations, in which it is assumed that the observation results are distributed
accordrepresented as follows.
is calculated ɣ (3).
ing to the normal law in accordance with the Gaussian function [
= ( 1,… ; к , 1,. . , к ; ∑1, . . , ∑к), = ( | , ∑
) ∑ =1 ( | ,∑
) =∑ ∑ =1 ( − )( − ) = 1
= = ∑ =1 , > ∆.
The procedure is stopped if the difference of the hidden variables does not exceed the specified constant (5). This model is based on the methods of mathematical statistics.
= max{ | − 0 }.
Data mining software tools enable using and compare different approaches to implement and choose the most appropriate method. 3
Re s ults
Тo test the clustering methods, information was prepared with the assignments of the
initial data: input, descriptive and output data. The following features were selected as
input data: gender, age, marital status, income, store category. For informative purposes
full name and purchase amount were used. The rest of the data is highlighted as not
being used. For the analysis, the Deductor Studio platform[
Segmentation analysis was performed using three k-means methods, Kohonen maps and EM clustering. When using the k-means method, the initial data was assigned as above was said: input, informative and output data. The choice of the number k can be based on theoretical justification or intuition. The number of clusters was set equal to 5. Result cluster profiles are shown in Figure 2. They display statistical information on clusters as a percentage. The data “category of stores” (this column in the table in the Figure 1 is not shown) includes such stores as grocery, construction, furniture, household, pharmacies, etc., a total of 51 records. Figure 2 shows this category in the highlighted part of the table on the right. In total, five clusters were identified by categories: shops, marital status, gender, income and age.
The main difference between Kohonen's self-organizing maps and the k-means method is that in it all neurons (class center and node) are ordered into some structure. Let's continue the example with clients for certain preferences. The same document for processing was selected. The same input and information data were leaved. The results of the execution of the algorithmare displayed on maps and a separate map is built for each input parameter. The convenience of the map (Figure 3) is that a user can click on the cluster number and see the location of the cluster on other maps that were built using the input values. For example, cluster 0 includes men, with an average age of 36 years, with marital status “married”, with an income of 50 thousand, and chose the following categories of stores: construction, food.
clasters links profiles age significance sex the shops
For EM clustering, the same sample, leaving the same input data and settings were chosen too. Clustering with this method allows user to choose automatic or fixed definition of clusters at the step of setting parameters. According to the connection of the clusters, it can be said, that the systemhas allocated the maximum number of clusters that were set.
Claster links map claster
What if Profiles claster age family
cash income Clistance matrix the clasters
The shops grocery store hardware beaty shop clothing furniture auyo shop pharmacy
The choice of the method of initial initialization was made. If it is chosen at random, we get 5 clusters. We can see that cluster 1 includes both men and women with the status of married / married or widower aged 23 to 65, including all categories of stores (Figure 4). For example, cluster # 2 included 70% of women and 30% of men aged 16 to 21, single and unmarried, interested in such categories of stores as food, beauty, pharmacy, clothing, household.
Let's configure for a fixed number of clusters. The choice of the method of initial initialization of clusters is given. Let's choose it at random, we get 5 clusters. We can see that cluster 1 includes both men and women with the status of married / married or widower aged 23 to 65, including all categories of stores (Fig ure 4).
For example, cluster 2 included 70% of women and 30% of men aged 16 to 21, single and unmarried, interested in s uch categories of stores as food, beauty, pharmacy, clothing, household. When the method of initial initialization of clusters is chosen “from the training set”, we get 5 clusters (Figure 4). In this case, cluster 2 includes unmarried women from 16 to 20 years old who are interested in the categories of the store: food, clothing, household and pharmacy.
Claster lincs x whats if x clastrr matrix x profiles claster x table
clasyers atAttrtibriubtuete Cash income
significance family significance age
significance sex significance the shop significance Namber of records 12 signatificance married not married widower married frequency Considered algorithms have fast cluster detection speed. However, the simplest for the user the clustering algorithmby Kohonen Maps became. In this algorithm, isn’t needed to specify the number of clusters at the input, the number of clusters determining is provided by the method. Also, in addition, this method has its own special display method, in the form of colored maps. On the maps, it is convenient immediately to determine where which cluster is located and what data is included in this cluster. Also, this algorithmhas good resistance to noisy data. Thus, with the help of Kohonen maps, the analyst can obtain more accurate segmentation results. 5
To choose the method on the context of solving market segmentation problemcomparative analysis of three clustering methods was performed. K-means, ЕМ-clustering method, and Kokhonen’s neural networks method were considered and analyzed applied for market segmentation. Kokhonen’s neural networks method was recommended as preferable because accuracy is the most critical requirement for the subject areas linked with marketing sphere. Databases information about customers gives opportunity to group clients by different preferences and take interests these groups into account when making decisions in marketing to improve business processes.