Data Analytics: Exploiting the Data Warehouse - Data Warehouse Systems: Design and Implementation - page 349

[NWDW Decision Tree].[Permanent Employees] =
T.[PermanentEmployees] AND
[NWDW Decision Tree].[Year Established] = T.[YearEstablished] AND
[NWDW Decision Tree].[Store Surface] = T.[StoreSurface] AND
[NWDW Decision Tree].[Parking Surface] = T.[ParkingSurface]
WHERE [High Value Cust] = 1

This query results in the following table, where values in the column Prob=1
indicate the probability of being a high-valued customer:

CompanyName BusinessType PE YE AP SS PS Prob=1 Prob=0

L ' Amour Fou Restaurant 4 1955 2 1178 918 0.7551 0.2448

Le Tavernier Pub 1 1984 1 2787 438 0.5326 0.4673

Potemkine Restaurant 5 1956 1 773 460 0.7551 0.2448

Flamingo Restaurant 3 1960 2 2935 1191 0.6041 0.3958

Pure Bar Pub 3 1989 2 1360 307 0.5326 0.4673

··· ··· ··· ··· ··· ··· ··· ··· ···

Clustering

We will now show how to build a clustering model to find out a customer
profile structure, using the view TargetCustomers and the parameters depicted
in the right-hand side of Fig. 9.7 . Then, given the table of prospective
customers ( NewCustomers ), we can predict to which profile each new customer
is likely to belong. Figure 9.9 shows the result of the clustering algorithm. The
shadow and thickness of the lines linking the clusters indicate the strength
of the relationship between the clusters, the darker and thicker the line, the
stronger the link between two clusters. The profiles of some of the clusters
are given in Fig. 9.10 . These profiles indicate, for example, the number of
elements in each cluster and the distribution of the attribute values within
each cluster. We can see, for example, that Cluster 5 contains few high-valued
customers.
In clustering models, a content query asks for details about the clusters
that were found. A prediction query may ask to which cluster a new data
point is most likely to belong.
Once the model is built (in this case, called NWDW Clustering ), we can find
out the characteristics of the clusters produced. Since in Analysis Services, the
clustering structure is a tree such that below the root ( NODE TYPE =1) there
is a collection of flat nodes (i.e., NODE TYPE =5). Thus, since all clusters
have a node type of 5, we can easily retrieve a list of the clusters by querying
the model content for only the nodes of that type. We can also filter the
nodes by support. The query shown below displays the identifier, the name,
the support (the number of elements in the cluster), and the description (the