Practical Problem Solving - Java Data Mining: Strategy, Standard, and Practice - page 322

71. lDataSetName,

72. lModelName,

73. lApplyDSSettName,

74. getVendorDatasetURI(lApplyOutTableName));

75. lVerifTask lApplyTask.verify();

76. if (lVerifTask ! null) {

77. reportError(lVerifTask.getReportText());

78. return;

79. }

80. String lApplyTaskName lModelName "_ApplyT";

81. mJDMConnection.saveObject(lApplyTaskName,

82. lApplyTask,

83. true);

84. lSuccess executeTask(lApplyTaskName);

85. if (!lSuccess) {

86. return;

87. }

88. }

89. }

As can be seen at line 60, we specify that we want to generate the
cluster identifiers into an attribute with a name obtained with the
model name prefix concatenated with the maximum number of clus-
ters for this model. After this method is executed, we have clusters
associated with the customers inserted into specific table columns.
Once each customer in the database is associated with clusters
from the different clustering models, the next step is to decide which
clustering model should be kept. The technique generally used to
compare clustering models is called profiling . This consists of com-
paring the distributions of some attributes for customers belonging
to each cluster with the distribution of the same attributes for the
customers not belonging to this cluster, that is, the distribution
between attributes values for cases assigned to a given cluster and
the reaming cases is far apart). With this analysis, the marketer
decides to name a given cluster with the attribute names and values
that best describe the cluster.
For example, an attribute describing “work class” could be pro-
filed for a specific cluster to see if this attribute has a distribution
very different from the profile obtained on the population not
belonging to this cluster. Profiling can be done on the attributes used
to build the clustering model as well as other attributes. In our exam-
ple, the clusters are made out of the psychographics data provided
by the service bureau, but the marketer could very well profile the
different clusters on attributes representing behavioral information,
such as how customers buy the product.