Propose how a separate memory-based reasoning analysis could provide additional insight into the problem.

Background information

Decision has been made

Apex Decision Support, Inc., can provide important insights into what
the data mining techniques might yield instead of having to plan and
implement an entire project to gather this information.

Previous problem

Classifying warranty customers. Many people who have purchased the
platinum warranty for their cars are unhappy with this service option.

Customer data (including age, address, date of purchase, and model
purchased) for the 540,000 purchasers of the platinum warranty.

In addition, you have warranty satisfaction levels for the 295,000
people who sent back the customer response cards.

We have Retro's automotive service data (type, cost, length of time for
service or repair).

We have lots of data about these platinum customers, but what really
separates the satisfied ones from the unsatisfied ones?

Next, problem involves classifying customer markets.

Recieved an e-mail message from William Bellman, the marketing director
at Retro Motors. Bill has some concerns about an advertising piece that
ran in newspapers across the United States.

Advertising piece didn't generate the expected results, and many people
are upset about the cost of the piece.

Alissa Quinn, your data mining consultant from Apex, indicates that
different data mining techniques may be useful for Retro's problems.

She suggests that using a decision tree analysis would be useful for the
warranty service problem classifications.

Alissa also suggests that neural networks and memory-based reasoning
would be useful for the customer market problem.

Alissa used the techniques to provide some simple outputs from each of
the models. They will serve as input to consider how the techniques work
and how data mining might address each problem area.

Data and Email results

Sources of data to address the warranty problem facing Retro:

Platinum customer demographics

Customer response cards

Service history

The following data are available from customer demographics:

Field Values

Vehicle ID # Unique identifier for each vehicle

Last name Text

First name Text

City Text

State Text

Zip code Number

Age Number

Gender [Male], [Female]





The following data are available from customer response cards and
related e-mails:

Field Values

Vehicle ID# Unique identifier for each vehicle

Satisfaction level [Very Low], [Low], [Fair], [High], [Very High]

Comments Text

Date Date

The following data are available from the service history:

Field Values

Vehicle ID# Unique identifier for each vehicle

Type of repair [Oil change], [Air conditioning service], [Transmission],
[Engine], [Other]

Date of service Date

Duration of service Hours



From; Alissa Quinn, Apex Decision Support

Subject: Neural Network and Memory-based Reasoning Analysis

Received the data that you classified as coming from customer
demographics, customer response cards, and vehicle service histories. I
believe that the decision tree technique is the best choice for
analyzing this data.

Using a tool to perform a decision tree analysis. Rather than
implementing an algorithm for selecting the attributes on which to
split, used a simple decision tree tool to select the order of
splitting. Then, the tool performed binary splits for each attribute.
This required that I combine the groups of data that you sent me and
transform the joined data as follows:

Field Value

Vehicle ID# Unique identifier for each vehicle

Satisfaction High [Yes], [No]

Age > 40 [Yes], [No]

Gender = Female [Yes], [No]

Minor Repair [Yes], [No]

Duration > 4 Hours [Yes], [No]

Duration > 8 Hours [Yes], [No]

Keep in mind that this transformation is a requirement of the
application, and not necessarily a requirement of decision trees. For
your current data, satisfaction was considered high if the customer
responded "High" or "Very High," and a repair was considered minor if it
was an "Oil change" or "Air conditioning service."

Here is a screen capture of the decision tree for you to use in your
analysis:

Here is a screen capture of the decision tree for you to use in your
analysis:



Customer Market Problem Details

Because you are interested in developing a model of customer
satisfaction levels, you want to use the customer response data. To
utilize the other data, you need to join the records of the other data
sources. If the data are stored in a database system, this is
accomplished by joining the tables. For example, you can join the
customer demographic information to the response data by using the
vehicle ID#. Likewise, you can join the service data by using the
vehicle ID# and the date of service. The date of service is an important
field to use because you want to determine whether a causal relationship
exists between the type and duration of service and the level of
satisfaction

From: William Belman

Subject: Data Mining Help

At our monthly sales meeting, Retro sales managers were discussing the
success, or lack of success, of an expensive newspaper insert that we
produced for our dealers to use as advertising in their local papers.
Many dealers complained that the cost was high and the response was low.
I promised my sales managers that I would investigate the problem. We
can't afford to spend the kind of time and energy that went into that
advertising piece if the effort is wasted.

I heard that your department is using data mining to gain greater
insight into Retro's business problems. We really need to understand why
our advertising piece wasn't as effective as intended. In addition, we
need to determine how to better predict the effectiveness of this type
of insert ad in the future.

Although the dealers, because they are semiautonomous, did not have to
use the insert, we encouraged them to do so. Initial results from the
first dealers were not encouraging, and only about half of them made
marginal profits sufficient to justify the cost of producing and
distributing the insert. Perhaps we can figure out where the insert has
been effective and limit our use to those areas. For example, when
dealers from those areas request the insert, we could steer them away
from it if we don't think it will be effective. I really need some
analysis about this. Can you help?

From: Alissa Quinn, Apex Decision Support

Subject: Neural Network and Memory –based Reasoning analysis

We have completed the neural network analysis of the advertising data
that you provided us with last week. Attached you will find a complete
report of the results. Per your request, we analyzed the data against
the list of cities being considered for running the advertisement next
month. You will see from the report that Fargo, North Dakota, appears to
be a good candidate for the advertisement, while Detroit, Michigan, does
not.

Although the results seem encouraging for several of the cities, you may
want to consider further analysis of the data using memory-based
reasoning. This technique may further corroborate the results of the
neural network analysis, as well as provide some explanation for why
certain cities appear to be better candidates than others.



Neural Network Analysis

Customer: Retro Motors

Project #: 2000-01865

Data

One thousand records were used for this analysis. The data was roughly
split between Wichita and Phoenix. From the information provided,
Phoenix, Arizona, was considered a bad result, whereas Wichita, Kansas,
was a good result. The following data fields were selected for analysis:

Data Field Values

Age [16-25], [26-35], [36-45], [> 45]

Gender [Male], [Female]

Education level [High School], [Some College], [Undergraduate Degree],
[Some Graduate],[Graduate Degree], [Other]

Number of children [0], [1-2], [3], [> 4]

Housing [Own home], [Rent], [Other]

Number of vehicles [0], [1], [2], [> 2]

Current Retro customer [Y], [N]

Income level [< 15,000], [15,000 - 25,000], [25,000 - 40,000], [>
40,000]

Promotional gift selected [Hat], [Shirt], [Movie Tickets], [None]

Comments:

Customers who selected a promotional gift were considered respondents to
the advertisement, because the advertisement tells the customers to ask
for their gift.

Neural Network

Training Set Size: 1,000

Iterations until convergence: 7

Because Wichita was considered a successful placement of the
advertisement and Phoenix was not, the output of the neural network is a
value predicting whether a different city (for example, Akron, Ohio)
will be a good or bad candidate for future advertising campaigns.

Results

Input Description: We utilized the most current demographic data for the
cities that Retro requested. The source of the data was the most current
U.S. census information.

Input Size: We used 250 random samples from each prospective city.

Method: Values from the input were fed into the neural network. Each
prediction was averaged, resulting in a gross prediction for the city.

Output Description: Output values range from 0 to 1. As the output
approaches 0, the city being tested is more like Phoenix (poor). As the
output nears 1, the city being tested is more like Wichita (excellent).

Output Accuracy: + or - 1%

Figure 1: Neural Network Output for Newspaper Advertising Success in
Given City