Java Reference

In-Depth Information

values. As we discuss in Sections 4.2 and 4.3, classification and

regression also differ in how model quality is assessed.

Unsupervised functions do not use a target, and are typically

used to find the intrinsic structure, relations, or affinities in a dataset.

Unlike supervised mining functions, which predict an outcome,

unsupervised mining functions cover a wide range of analytical

capabilities including clustering and association. Clustering may be

used to identify naturally occurring groups in the data, for example,

similar proteins or cancer cells, or retail customer segments. Associ-

ation models return items and rules that can be used, for example,

to identify products for cross-sell to retail customers. There are

other unsupervised mining functions, such as sequential patterns

[Wang

2005], feature extraction [Lui

1998], and anomaly detec-

tion [Margineantu

2005], not yet officially covered by JDM.

Another dimension of data mining involves whether the result-

ing model is
descriptive
or
predictive.
Descriptive data mining

describes a dataset in a concise, enlightening, and summary manner,

and presents interesting generalized properties of the data. Descriptive

data mining results in models that provide
transparency
, that is, the

ability to understand why the model behaves as it does. The extent to

which a model is descriptive often depends on the algorithm used to

produce it. For example, in supervised learning, decision trees typi-

cally provide human interpretable rules that explain why a given

prediction was made, whereas a neural network used on the same

data provides no readily discernable understanding, and are consid-

ered “black-box” or
opaque
. Most unsupervised mining functions,

such as clustering or association rules, by definition are descriptive.

Mining functions and algorithms can be both descriptive
and
predic-

tive, such as the decision tree algorithm.

Predictive mining functions perform inference on the available

data, and attempt to predict outcomes or assignments for new data.

In addition to the prediction, predictive mining functions may also

provide a probability or confidence as to how strong the prediction

is based on what the model knows. For example, a clustering model

may assign a case to Cluster 5 with a 95 percent probability (a

strong assignment), or a classification model may predict the

customer will churn with a 55 percent probability (a weak predic-

tion). Supervised mining functions by definition are predictive.

Supervised algorithms supporting predictive data mining include

naïve bayes, neural networks, support vector machine, and deci-

sion trees. The clustering mining function, with algorithms such as

k-means, may also be considered predictive when used to assign

Search WWH ::

Custom Search