Data Mining Process - Java Data Mining: Strategy, Standard, and Practice - page 56

settings. JDM also provides algorithm level settings, where a user can
select a specific algorithm and adjust settings manually. JDM also
provides detailed attribute-level specifications, called logical data , for
defining how the input physical data should be interpreted. For exam-
ple, in Section 3.1.2, we gave an example involving whether to inter-
pret a number attribute as continuous (age) or discrete (rating). This
can be specified as input to model building using the logical data
specification.
For model assessment, JDM provides capabilities to test supervised
models using a variety of techniques, as well as to inspect model
details such as cluster definitions, or the rules associated with a deci-
sion tree. Specifically, JDM provides interfaces for assessing classifica-
tion models via the confusion matrix, lift, and receiver operator
characteristics (ROC); regression models via error metrics; clustering
models by viewing cluster rules and centroids; association models by
filtering and inspecting rules; and attribute importance models by
viewing the ranking of attributes. Models produced using specific
algorithms may also have corresponding model details to provide
greater insight into the results produced. These concepts are discussed
in detail in Chapters 7, 8, and 9.

3.1.5 Evaluation Phase

Whereas building and testing models is the fun part, the next phase,
evaluation , cannot be overlooked. Before unleashing a model in a
business application or process, we need to assess how well it meets
the business objectives set out in the business understanding phase.
Although we may have high quality models from the modeling
phase, they may still not satisfy the business objectives. For example,
an exploratory model may produce superior results using attributes
fully populated with values in the data sample, but that are not pop-
ulated for most customers in practice. During the evaluation phase,
we review the steps leading up to the model and its quality assess-
ment to determine if some aspect of the business problem has not
been addressed, or not addressed adequately. The objective for this
phase is to decide whether or not the model can be deployed in the
business application or process.
As noted for the modeling phase, JDM provides much of the raw
information needed to support the evaluation phase, in terms of test
metrics and model details. The evaluation phase relies on domain
knowledge and critical thinking to assess whether the data mining
models will address the business need.