244 Part II • Information Technology
Case Worker Advisor incorporates the cultural aspects and
philosophy of the Navajo case workers, streamlines the pre-
vious assessment methods, and assists less experienced
case workers. “In programming the knowledge automation
system we accounted for our unique cultural heritage, while
following complex federal, state, and tribal guidelines,”
according to the Navajo Nation’s TANF project director
(Exsys Inc., 2010).
Neural Networks
Whereas expert systems try to capture the expertise of
humans in a computer program, neural networks attempt to
tease out meaningful patterns from vast amounts of data.
Neural networks can recognize patterns too obscure for
humans to detect, and they adapt as new information is
received.
The key characteristic of neural networks is that they
learn.The neural network program is originally given a set
of data consisting of many variables associated with a
large number of cases, or events, in which the outcomes
are known. The program analyzes the data, works out all
the correlations, and then selects a set of variables that are
strongly correlated with particular known outcomes as the
initial pattern. This initial pattern is used to try to predict
the outcomes of the various cases, and these predicted
results are compared to the known results. Based on this
comparison, the program changes the pattern by adjusting
the weights given to the variables or by changing the
variables. The neural network program then repeats this
process over and over, continuously adjusting the pattern
in an attempt to improve its predictive ability. When no
further improvement is possible from this iterative
approach, the program is ready to make predictions for
future cases.
This is not the end of the story. As more cases
become available, these data are also fed into the neural
network, and the pattern is once again adjusted. The neural
network learns more about cause-and-effect patterns from
this additional data, and its predictive ability usually
improves accordingly.
Commercial neural network programs (actually,
these are shells) are available for a reasonable price, but
the difficult part of building a neural network application is
data collection and data maintenance. Still, a growing
number of applications are being deployed. Neural net-
works are typically used either to predict or categorize, but
to do so in an inductive manner rather than deductively.
Table 6.2 lists examples of current uses of neural networks.
Let us consider some neural network examples.
American Express uses a neural system to read handwrit-
ing on credit card slips. The state of Wyoming uses a
neural system to read hand-printed numbers on tax forms.
Oil giants such as Arco are using neural networks to help
pinpoint oil and gas deposits below the earth’s surface.
Neural networks are being used to predict the total contin-
gency costs on construction projects. In one especially
interesting application, a neural network was designed to
predict the expected revenue range of a movie prior to its
theatrical release—and the neural network resulted in a
much better prediction than other statistical methods
currently employed (NeuroDimension, 2010).
Many of the larger banks in the United States use
neural networks to develop credit scores for consumers
and small and medium businesses. Then the credit scores
become the basis for approving loans or deciding on a
collection strategy. Mellon Bank installed a neural network
credit card fraud detection system. When a credit card is
swiped through the card reader in a store, the transaction is
sent to Mellon’s neural system. By analyzing the type of
transaction, the amount spent, the time of day, and other
data, the neural network makes a fraud prediction in a few
seconds and either approves or denies the transaction or
feeds the predictive score to a human analyst who makesTABLE 6.2 Uses of Neural NetworksCategorization Prediction/Forecasting
Credit rating and risk assessment Share price forecast
Insurance risk evaluation Commodity price forecast
Fraud detection Economic indicator predictions
Insider trading detection Process control
Direct mail profiling Weather prediction
Machinery defect diagnosis Future drug performance
Character recognition Production requirements
Medical diagnosis
Bacteria identification