Answers to Exercises

Supplementary Readings, Exercises, and Assignments: All other readings,
exercises, and assignments are posted to our electronic course site. III. Course
Objectives and Learning Goals. This course will focus on the design and
management of data warehouse (DW) and business intelligence (BI) systems.
The DW is the ...

Part of the document


CHAPTER 1 Review Questions 1. Differentiate between the following terms:
a. A data warehouse stores historical data for purposes of decision
support. An operational database is designed to efficiently process
current transactions.
b. Training data is used to build a supervised learner model. Test data is
applied to test the accuracy of a created model.
c. An input attribute is used to help create a model that best
differentiates the values of one or more output attributes.
d. Shallow knowledge represents facts that can be found using simple
database queries. Hidden knowledge represents structure in data that is
usually not uncovered with simple database queries.
e. The exemplar view hypothesizes that we store examples of concepts to
help categorize unknown instances. The probabilistic view states that we
build and store generalizations of concept instances for purposes of
future categorization.
f. The classical view requires all concept-defining properties to be
present for an instance to be categorized with a specific concept. In
contrast, the probabilistic view does not have rigid requirements for
individual attribute values. Rather, concept definitions are stored as
generalizations. An instance is categorized with the concept whose
generalization most closely matches its attribute values.
g. Supervised learning requires instances to have one or more output
attributes. The purpose of a supervised learner model is to determine
correct values for the output attributes. With unsupervised clustering,
an output attribute is not present. The purpose of unsupervised
clustering is to find a best-fit partition of the data instances.
h. Relative to customer purchases, the actual value of a customer or client
is the amount of money spent by the client. The instrinsic value of the
client is the amount the client is expected to spend.
2. Choose whether supervised learning, unsupervised clustering or database
query is most suitable. As appropriate, state any intital hypotheses you
would like to test. List several attributes you believe to be relevant
for solving the problem.
a. Supervised learning is the likely choice where the output attribute is
returned to work. Possible input attributes include age, current pain
level, income, occupation as well as several others.
b. Supervised learning is a reasonable choice where the output attribute
is whether a given vehicle has been involved in an accident. Input
attributes to consider include: vehicle type, tire type, mileage, the
plant where the vehicle was manufactured, the plant where the tires
where manufactured, when the vehicle or tire was manufactured (day of
week), road conditions at the time of the accident etc. An argument
for database query can also be made.
c. We first assume that each product is an attribute with two possible
values. For a given customer instance and product, the value is yes
if the product was purchased by the customer and no if the product was
not purchased. Unsupervised clustering will help determine which
products are most often purchased together. We may also consider
several supervised mining sessions where each product takes its turn
as an output attribute.
d. Database query
e. Unsupervised clustering is one choice. This assumes we are not trying
to determine the value of any one attribute. However, if we wish to
develop a model to determine a likely favorite spectator sport based
on a person's age, height, and weight, supervised learning is the best
choice.
3. Medical students learn surgical procedures by observing and assisting
trained doctors. As individual observations lead to generalizations about
how to perform specific operations, the learning is inductive. 4. As the web site contains a wealth of information about data mining,
answers will vary. 5. There are several possibilities.
a. Here are some choices: GPA, total number of earned credits, number of
years in school, average credits taken per semester, extra curricular
activities, and whether the person has a job.
b. One classical view definition is:
A GPA of 3.0 or higher, averages 12 or more credits per semester and
has a declared major.
c. A probabilistic definition can be stated as:
An above average GPA, usually carries an average to above average
credit load, and often times has a declared major.
d. An exemplar view definition lists several examples and non-examples of
good students. A new instance is given the classification associated with
the best matching exemplar.
6. Let's pick sore throat as the top-level node. The only possibilities are
yes and no. Instances one, three four, eight, and ten follow the yes
path. The no path shows instances 2,5,6,7 & 9. The path for sore throat =
yes has representatives from all three classes as does sore throat = no.
Next we follow the sore throat = yes path and choose headache. We need
only concern ourselves with instances 1,3,4, 8 & 10. For headache = yes
we have instances 1 (strep throat) ,8 (allergy ), & 10 (cold). For
headache = no we have instances 3 (cold) and 4 (strep throat).
Next follow headache = yes and choose congestion(the only remaining
attribute. All three instances show congestion = yes, therefore the tree
is unable to further differentiate the three instances. A similar problem
is seen by following headache = no. Therefore, the path following sore
throat = yes is unable to differentiate any of the five instances. The
problem repeats itself for the path sore throat = no. In general, any top-
level node choice of sore throat, congestion, or headache gives a similar
result.
CHAPTER 2
Review Questions 1. Differentiate between the terms:
a. A data mining strategy is a template for problem solving. A data mining
techique involves the application of a strategy to a specific set of
data.
b. A set if independent variables is used to build a model to determine the
values of one or more dependent variables.
2. Yes on both counts. As one example, feed-forward neural networks and
linear regression models can both be used for estimation problems.
Likewise, neural networks can be used for estimation as well as
prediction problems. It is the nature of the data, not the data mining
technique, that determines the data mining strategy. 3. Is each scenario a classification, estimation or prediction problem?
a. This is a prediction problem as we are trying to determine future
behavior.
b. This is a classification problem as we are classifying individuals as a
good or poor secured credit risks.
c. This is a prediction problem.
d. This is a classification problem as the violations have already
occurred.
e. This is a classification or estimation problem depending on how the
output variable(s) are represented.
4. There are no absolute answers for this question. Here are some
possibilities.
a. For 3a: As there is an output attribute and the attribute is numeric, a
neural network is a good choice. Statistical regression is also a
possibility.
For 3b, 3d, 3e: A decision tree model or a production rule generator
is a good choice as an output attribute exists and we are likely to be
interested in how the model reaches its conclusions.
For 3c: A neural network model is a best choice as the output can be
interpreted as the probability of a stock split.
b. For 3a: Any technique limited to categorical output attributes would be
of limited use as we are interested in a numeric output.
For 3b, 3d, 3e: Any technique that does not explain its behavior is a
poor choice.
For 3c: A numeric output between 0 and 1 inclusive that can be treated
as a probability of a stock split allows us to make a better
determination of whether a stock is likely to split. Therefore, any
technique whose output attribute must be categorical is a poor choice.
c. Answers will vary.
5. For supervised learning decision trees, production rules and association
rules provide information about the relationships seen between the input
and output attributes. Neural networks and regression models do a poor
job of explaining their behavior. Various approaches to unsupervised
clustering have not been discussed at this point. 6. As home mortgages represent secured credit, model A is likely to be the
best choice. However, if the lender's cost for carrying out a forclosure
is high, model B may be a better alternative. 7. As the cost of drilling for oil is very high, Model B is the best
choice. 8. If clusters that differentiate the values of the output attribute are
formed, the attributes are appropriate. 9. Each formed cluster is designated as a class. A subset of the instances
from each class are used to build a supervised learner model. The
remaining instances are used for testing the supervised model. The test
set accuracy of the supervised model will help determine if meaningful
clusters have been formed. Data Mining Questions 1. The architecture of the network should appear similar to the network
shown in Figure 2.2. However, the network should have 4 input-layer
nodes, 5 hidden-layer nodes, and 1 output-layer node. 2. Students find this to be an interesting exercise. You may wish to
discuss credit card billing categories and help students set up their
individual spreadsheets. Computational Questions 1. Consider the following three-class confusion matrix.
a. 86%
b. 48, 45, 7
c. 2
d. 0 2. Suppose we have two classes each with 100 instances.
a. 40
b. 10
3. Consider the confusion matrices shown below.
a. 2.008
b. 2.250
4. Let + represent the class of individuals respondi