CHAPTER 11 LEARNING FROM DATA

sciencediscussionΤεχνίτη Νοημοσύνη και Ρομποτική

20 Οκτ 2013 (πριν από 3 χρόνια και 8 μήνες)

107 εμφανίσεις

CHAPTER 11

LEARNING FROM DATA



TEST YOUR UNDERSTANDING


1.

Define a neural network in your own words. What does the technology

attempt to do?


Neural networks are a knowledge
-
based technology modeled after the human brain’s
network of electrically inter
-
conn
ected processing elements called neurons. It is a
learning
system that creates a model based on its inputs and outputs.
The technology
attempts to predict, classify, or cluster data sets
.


2.

“The interesting aspect of a neural net is its known contributions
in solving
cumbersome problems that traditional computers have found difficult to track.” Do
you agree? Why or why not?


The basic feature of neural networks is their ability to learn from examples. They do not
need to build analytical models or devise a s
et of rules to solve a problem. Therefore they
can more easily tackle complex problems.


3.

Explain how a neural network functions. Give an example of your own.


A neural network consists of neurons, connected by axons that receive inputs
from other
neurons.

A neuron evaluates the inputs, determines their weights, sums the weighted
input, and compares the total to a threshold. If the sum is greater than a certain threshold,
the neuron fires. Otherwise, it generates no signal.


4.

In your opinion, how severe are
the limitations of a neural network?


Neural networks are black box systems. They are unable to provide explanation. From a
business perspective, this is a severe limitation, because managers need to communicate
and justify the decisions they may take. Neu
ral networks do not provide such support.


5.

How are neural networks different from knowledge automation systems?


Neural networks differ from knowledge automation systems in the following aspects:



Knowledge automation systems learn by rules, and neural nets

learn by example.



Knowledge automation systems are sequential in nature and explain their decisions,
where neural nets do not have an explanatory facility.



Neural networks continue learning as the problem environment changes, and
knowledge automation syst
ems depend on complete data before offering a final
solution.


6.

Explain in some detail how neural networks learn.


CHAPTER 11

LEARNING FROM DATA



11
-
2

Learning occurs when the network changes weights and modifies its activity in response
to its inputs. Learning may be supervised or unsupervis
ed. In
supervised learnin
g, the
neural network needs a teacher, with a
training set
of examples of input and output. In
unsupervised
(also called
self
-
supervise
d)
learnin
g, no external factors influence the
adjustment of the input’s weights. The neural net
work has no indication of correct or
incorrect answers. It adjusts solely through direct confrontation with new experiences.


7.

How are inputs and outputs used to contribute to a solution?


The neural network needs a teacher, with a
training set
of examples

of input and output.
Each element in a training set is paired with an acceptable response. That is, the actual
output of a neural network is compared to the desired output. The network makes
successive passes through the examples, and the weights adjust t
oward the goal state.
When the weights represent the passes without error, the network has learned to associate
a set of input patterns with a specific output. This is more like learning by reinforcement.



KNOWLEDGE EXERCISES


1.

Search on the Internet or i
n related literature and write an essay detailing a neural

network application in business. What did you learn from this exercise?


Students are expected to search on the Web for a business case where neural networks
were used to find a solution.


2.

Suppose
you have a small kiosk database of purchased items.

The available food items
in the kiosk are as follows:

• Coca
-
Cola

• Pepsi
-
Cola

• Sprite

• Budweiser beer

• Guinness beer

• Estrella chips

• Pringles chips

• Taffel chips


The database contains the follow
ing purchase transactions:


TID





Items Bought


1


Coca
-
Cola, Budweiser beer, Pringles chips

2


Coca
-
Cola, Taffel chips

3


Budweiser beer, Pringles chips

4


Pepsi
-
Cola, Budweiser beer, Guinness beer, Estrella chips

5


Sprite, Estrella chips

6



Pepsi
-
Cola, Budweiser beer, Estrella chips.

CHAPTER 11

LEARNING FROM DATA



11
-
3

7


Sprite

8


Budweiser beer, Guinness beer, Estrella chips

9


Pepsi
-
Cola, Estrella chips

10


Coca
-
Cola, Pringles chips


What kind of rules do you get with a confidence threshold of 0.0 and a support
th
reshold of 0.2?



Create a hierarchy for the food items in the kiosk, and try to determine if you could get
more meaningful information by using multilevel association rules. Is it possible to set
different support thresholds on different hierarchy levels?


Source
: Han and Kamber 2000



The database contains 10 transactions

3 transactions contain Coca
-
Cola

3 transactions contain Pepsi
-
Cola

2 transactions contain Sprite

5 transactions contain Budweiser beer

2 transactions contain Guinness beer

5 transaction
s contain Estrella chips

3 transactions contain Pringles chips

1 transaction contains Taffel chips


1 transaction contains Coca
-
Cola and Budweiser beer

2 transactions contain Coca
-
Cola and Pringles chips

1 transaction contains Coca
-
Cola and Taffel chips

2
transactions contain Pepsi
-
Cola and Budweiser beer

1 transaction contains Pepsi
-
Cola and Guinness beer

3 transactions contain Pepsi
-
Cola and Estrella chips

1 transaction contains Sprite and Estrella chips

2 transactions contain Budweiser beer and Guinness
beer

3 transactions contain Budweiser beer and Estrella chips

2 transactions contain Budweiser beer and Pringles chips

2 transactions contain Guinness beer and Estrella chips



No rules can be found with a confidence threshold of 0.0 and a support threshol
d of 0.2.


Rules with confidence threshold of 0.0:

If a customer buys Coca
-
Cola, 100% she will not buy Pepsi Cola.

If a customer buys Guinness beer, 100% she will not buy Pringles chips.

Rules with support threshold of 0.2:

If a customer buys Pepsi, 66% sh
e will buy Budweiser beer. Support: 0.2

If a customer buys Budweiser beer, 66% she will buy Pringles chips. Support: 0.2

Meaningful rules:

If a customer buys Pepsi
-
Cola, 100% she will buy Estrella chips, this happened 30% of
the times.

CHAPTER 11

LEARNING FROM DATA



11
-
4

If a customer buys B
udweiser beer, 60% she will buy Estrella chips, this happened 30%
of the times.


3.

Using a realistic example, discuss in detail how neural networks can be used to solve
real life problems.


Students are required to look for a real life situation where neural

networks could be
used. They need to identify the business problem and imagine how neural networks could
be applied to solve the problem.


4.

An insurance company is planning to launch the sale of a newly developed insurance
product. As a promotional exercis
e, the company wishes to identify several potential
customers from the local area and send them details of this new product, together with
special discount offers. In the past, the company has conducted similar promotional
exercises, and the results were n
ot good. The promotion cost was high due to an
extremely low response rate. Very few people who responded to the promotions actually
purchased the products. This time, the company aims to reduce promotion cost and
have an increase in response rate. The com
pany possesses about 2,000 historical
records on the previous promotions. The data set stores details of the prospective
customers, such as age groups, income bands, occupations, and so forth. The records
can be categorized into “response with purchase,” “
response without purchase,” or
“no response.” If the company wants to use a classification technique to identify the
people who are likely to purchase the new product, describe the life cycle of such a
data
-
mining project.



The life cycle of the data min
ing project is as follows:



Data gathering:

the company here can depend on the historical data gathered from
the last promotion exercise.



Data preparation:

here the company modifies data in order to fit exploration
findings. The company evaluates the data q
uality, handles the missing data, and
quantifies it. At this step, the company develops a better understanding for some
variables and the irrelevance of others.



Model building:

once data is prepared, a model is built to explain patterns in data. In
our cas
e, we would like to find a pattern for those most likely to respond to the
promotion and purchase the product. Many techniques can be used: neural networks,
classification trees, association rules, etc.



Model testing and analysis of results:

at this stage

the model is reviewed thoroughly
to insure integrity, reliability, and usability.


5.

Review the neural network literature and cite one application of supervised learning
and another of self
-
supervised learning. What do you conclude is the difference
between

the two forms of learning?


Supervised learning application: Risk management, as a major bank will use a neural
network to appraise commercial loan applications. By training it with historical data
CHAPTER 11

LEARNING FROM DATA



11
-
5

about rejected and accepted applications, the net learned

how to identify risky loan
applicants who are most likely to default on their payment.


Self
-
supervised learning application: Forecasting fluctuations of the foreign exchange
market.


The main difference is that in the supervised learning the actual outpu
t is continuously
compared to the desired output and training is carried on until they match. In the case of
self
-
supervised learning, there is no favorable outcome, i.e.no correct or incorrect
answers, just to reach an answer.


6.

Identify whether each of t
he following applications is a candidate for knowledge
automation or a neural network:

a. Prediction of weather conditions,

b. Diagnosis of a diabetic condition by a specialist,

c. Verification of check signature, and

d. Recognition of characters on an inv
oice.


a)

Prediction of weather condition: Neural Networks

b)

Diagnosis of a diabetic condition by a specialist: Neural Networks

c)

Verification of check signature: Knowledge automation

d)

Recognition of characters on an invoice: Neural Networks


7.

You have a set of six

alphabetic characters: A, S, T, U, T, and E. Like a Scrabble game,
your job is to generate as many words as possible from these characters. If you were to
do it manually, you would have to think of a word, verify it with the dictionary, and
move on to the

next word. How would a conventional computer go through the same
process? Is this kind of job a candidate for a neural network?


The conventional computer will use probabilities to solve such a problem, that is listing
all combinations of these characters

and compare them to a built
-
in dictionary, until all
possible words are formulated.


This is not a job for Neural Networks. Conventional computers are more powerful at
finding all combinations of a set of letters and then matching them with a dictionary.


8.

If the company chooses to use the decision tree induction technique to build a
classification model, describe the main activities that need to be performed during the
project development.


The main activities are:

1.

Assign attributes to the nodes: starting

with the root node.

2.

Decide when a classification tree is complete: determining the ending nodes and the
desired attributes.

CHAPTER 11

LEARNING FROM DATA



11
-
6

3.

Assign classes to ending nodes: for example in Question # 5, the ending nodes can be
classified into three classes: “response with
purchase”, “response without purchase”,
and “no response”.

4.

Convert the classification tree into a set of rules, these rules will be used to classify
new candidates into the assigned classes.