Vous êtes sur la page 1sur 12

Journal of Theoretical and Applied Information Technology

2005 - 2009 JATIT. All rights reserved.

www.jatit.org

NEURAL NETWORKS IN DATA MINING


1 2
DR. YASHPAL SINGH, ALOK SINGH CHAUHAN
1
Reader, Bundelkhand Institute of Engineering & Technology, Jhansi, India

Lecturer, United Institute of Management, Allahabad, India E-mail: yash_biet@yahoo.co.in ,


alok_sc@yahoo.co.in

ABSTRACT

Companies have been collecting data for decades, building massive data warehouses in which to store it.
Even though this data is available, very few companies have been able to realize the actual value stored in
it. The question these companies are asking is how to extract this value. The answer is Data mining.

There are many technologies available to data mining practitioners, including Artificial Neural Networks,
Regression, and Decision Trees. Many practitioners are wary of Neural Networks due to their black box
nature, even though they have proven themselves in many situations. This paper is an overview of artificial
neural networks and questions their position as a preferred tool by data mining practitioners.

Keywords: Artificial Neural Network (ANN), neural network topology, Data mining, back propagation
algorithm, Advantages.
networks.
1. INTRODUCTION:

Data mining is the term used to describe the process


of extracting value from a database. A data-
warehouse is a location where information is stored.
The type of data stored depends largely on the type of
industry and the company. Many companies store
every piece of data they have collected, while others
are more ruthless in what they deem to be
important.

Consider the following example of a financial


institution failing to utilize their data-warehouse.
Another example of where this institution has failed
to utilize its data-warehouse is in cross-selling
insurance products (e.g. home, life and motor vehicle
insurance). By using transaction information they
may have the ability to determine if a customer is
making payments to another insurance broker. This
would enable the institution to select prospects for
their insurance products. These are simple examples
of what could be achieved using data mining.

Four things are required to data-mine effectively:


high-quality data, the right data, an adequate
sample size and the right tool. There are many tools
available to a data mining practitioner. These include
decision trees, various types of regression and neural
Income is a very important socio-economic indicator. artificial neurons and
If a bank knows a persons income, they can offer a An artificial neural processes information
higher credit card limit or determine if they are likely network (ANN), often using a connectionist
to want information on a home loan or managed just called a "neural approach to
investments. Even though this financial institution network" (NN), is a computation. In most
had the ability to determine a customers income in mathematical model or cases an ANN is an
two ways, from their credit card application, or computational model adaptive system that
through regular direct deposits into their bank based on biological changes its structure
account, they did not extract and utilize this neural networks, in other based on external or
information. words, is an emulation internal information that
of biological neural flows through the
system. It consists of an network during the
2. ARTIFICIAL NEURAL NETWORKS: interconnected group of learning phase.

37
Journal of Theoretical and Applied Information Technology

2005 - 2009 JATIT. All rights reserved.

www.jatit.org
of) units, but no priori
feedback knowledge.
connections are Another way is
present, that is, to 'train' the
connections neural network
extending from by feeding it
outputs of units to teaching patterns
inputs of units in and letting it
the same layer or change its
previous layers. weights
Recurrent according to
network: some learning
Recurrent neural rule. We can
networks that do categorize the
contain feedback learning
connections. situations as
Contrary to follows:
feedforward
networks, Supervised
recurrent neural learning or
networks (RNs) Associative
are models with learning in which
bi-directional data the network is
flow. While a trained by
feedforward providing it with
network input and
propagates data matching output
linearly from patterns. These
input to output, input-output
2.1 Neural RNs also pairs can be
Network propagate data provided by an
Topologies: from later
processing stages
Feedforward to earlier stages.
neural network:
The feedforward 2.2 Training Of 38
neural network Artificial Neural
was the first and Networks:
arguably simplest
type of artificial
A neural network
neural network
has to be
devised. In this
configured such
network, the
that the
information moves
application of a
in only one
set of inputs
direction, forward,
produces (either
from the input
'direct' or via a
nodes, through the
relaxation
hidden nodes (if
process) the
any) and to the
desired set of
output nodes.
outputs. Various
There are no
methods to set the
cycles or loops in
strengths of the
the network. The
connections exist.
data processing
One way is to set
can extend over
the weights
multiple (layers
explicitly, using a
develop its own In more practical
representation of terms neural
the input stimuli. networks are
non-linear
Reinforcement statistical data
Learning This modeling tools.
external teacher, or type of learning They can be used
by the system may be considered to model
which contains the as an intermediate complex
neural network form of the above relationships
(self-supervised). two types of between inputs
learning. Here the and outputs or to
Unsupervised learning machine find patterns in
learning or Self- does some action data. Using
organization in on the neural networks
which an (output) environment and as a tool, data
unit is trained to gets a feedback warehousing
respond to clusters response from the firms are
of pattern within environment. The harvesting
the input. In this learning system information from
paradigm the grades its action datasets in the
system is supposed good (rewarding) process known as
to discover or bad data mining. The
statistically salient (punishable) difference
features of the based on the between these
input population. environmental data warehouses
Unlike the response and and ordinary
supervised accordingly databases is that
learning paradigm, adjusts its there is actual
there is no a priori parameters. anipulation and
set of categories cross-fertilization
into which the 3. NEURAL of the data
patterns are to be NETWORKS IN helping users
classified; rather DATA MINING: makes more
the system must informed
decisions.
Journal of Theoretical and Applied Information Technology

2005 - 2009 JATIT. All rights reserved.

www.jatit.org
noise from mining provides
measurement information to
data; to control understand what
ill-defined is happening
problems; in inside the data
summary, to without a
estimate predetermined
sampled idea. Predictive
functions when data mining
we do not know allows the user
the form of the to submit
functions. It is records with
precisely these unknown field
two abilities values, and the
(pattern system will
recognition and guess the
function unknown values
estimation) based on
which make previous
artificial neural patterns
networks discovered form
(ANN) so the database.
prevalent a Data mining
utility in data models can be
mining. As data categorized
sets grow to according to the
massive sizes, tasks they
the need for perform:
automated Classification
processing and
becomes clear.
With their
model-free
estimators and
their dual
nature, neural
networks serve
Neural networks data mining in a
essentially myriad of ways.
comprise three
pieces: the Data mining is
architecture or the business of
model; the answering
learning algorithm;questions that
and the activation youve not
functions. Neural asked yet. Data
networks are mining reaches
programmed or deep into
trained to . . . databases. Data
store, recognize, mining tasks
and associatively can be classified
retrieve patterns or into two
database entries; to categories:
solve Descriptive and
combinatorial predictive data
optimization mining.
problems; to filter Descriptive data
clustering. The forecasting to
major difference the data mining
being that no industry.
groups have
been predefined. Financial
Prediction is the forecasting is of
construction and considerable
use of a model practical
to assess the interest. Due to
class of an neural networks
Prediction,
unlabeled object can mine
Clustering,
or to assess the valuable
Association Rules.
value or value information
Classification and
ranges of a from a mass of
prediction is a
given object is history
predictive model,
likely to have. information and
but clustering and
The next be efficiently
association rules
application is used in financial
are descriptive
forecasting. areas, so the
models.
This is different applications of
from predictions neural networks
The most common because it
to financial
action in data estimates the forecasting have
mining is future value of been very
classification. It continuous
popular over the
recognizes patternsvariables based
last few years.
that describe the on patterns
Some researches
group to which an within the data.
show that neural
item belongs. It Neural
networks
does this by networks, performed better
examining existing depending on
than
items that already the architecture,
conventional
have been provide statistical
classified and associations, approaches in
inferring a set of classifications,
financial
rules. Similar to clusters,
forecasting and
classification is prediction and
39
Journal of Theoretical and Applied Information Technology

2005 - 2009 JATIT. All rights reserved.

www.jatit.org
and output layer. feed-forward
are an excellent In each layer neural network.
data mining tool. there are one or
In data more processing
warehouses, neural elements (PEs).
networks are just PEs is meant to the previous
one of the tools simulate the
neurons in the layer. There are
used in data connections
mining. ANNs are brain and this is between the PEs
used to find why they are
in each layer
patterns in the data often referred to that have a
and to infer rules as neurons or weight
from them. Neural nodes. A PE (parameter)
networks are receives inputs
associated with
useful in providing from either the them. This
outside world or
weight is
3.1. Feedforward adjusted during
Neural Network: information on training.
associations, Information
One of the classifications, only travels in
simplest feed clusters, and the forward
forward neural forecasting. The direction
networks (FFNN), back through the
such as in Figure, propagation network - there
consists of three algorithm are no feedback
layers: an input performs loops.
layer, hidden layer learning on a
training The network. back
a FFNN predicted Back propagati
is as output is propagat on has
follows: subtracte ion is a finished,
d from learning the
Input the algorith forward
data is actual m for process
presente output adjusting starts
d to the and an the again,
network error weights. and this
and value for It starts cycle is
propagat the with the continue
ed networks weights d until
through is between the error
the calculate the between
network d. output predicted
until it The layer and
reaches neural PEs and actual
the network the last outputs
output then uses hidden is
layer. supervis layer minimize
This ed PEs and d.
forward learning, works
process which in backwar
produces most ds
a cases is through
The predicted back the
simplified output. propagat network.
process for ion, to
train the Once
error, is layers, and the
a and send outputs training
common their we want data.
method signals the
of forward network Summar
teaching , and to y of the
artificial then the compute, techniqu
neural errors and then e:
networks are the error
how to propagat (differen Present a
perform ed ce training
a given backwar between sample
task.The ds. The actual to the
back back and neural
propagat propagat expected network.
ion ion results) Compare
algorith algorith is the
m is m uses calculate network'
used in supervis d. The s output
3.2. The layered ed idea of to the
Back feed- learning, the back desired
Propagati forward which propagat output
on ANNs. means ion from that
Algorithm This that we algorith sample.
: means provide m is to Calculate
Backprop that the the reduce the error
agation , artificial algorith this in each
or neurons m with error, output
propagati are example until the neuron.
on of organize s of the ANN
d in inputs learns
40
Journal of Theoretical and Applied Information Technology

2005 - 2009 JATIT. All rights reserved.

www.jatit.org
continued Identifying tax
For each neuron, Update the fraud
calculate what the weights in the
output should have network Enhancing
been, and a end auditing by
scaling factor, until all finding
how much lower examples irregularities
or higher the classified
output must be correctly or Finance
adjusted to match stopping
the desired output. criterion Signature and
This is the local satisfied bank note
error. return(network)
verification Risk
Management
Adjust the weights Assign "blame"
of each neuron to for the local Foreign
lower the local error to neurons
exchange rate
error. at the previous
forecasting
Actual level, giving Bankruptcy
Algorithm: greater prediction
1. Initialize the responsibility to Customer credit
weights in the neurons scoring
network (often connected by
randomly) stronger Credit card
repeat weights. approval and
for each example e fraud detection
in the training set Repeat the steps Forecasting
do above on the economic
O = neural-net- neurons at the turning points
output(network, previous level, Bond rating and
e) ; forward pass using each one's trading
"blame" as its Loan approvals
T = teacher output error.. Economic and
for e financial
Calculate error (T fraud detection, forecasting
- O) at the output telecommunicati Marketing
units ons, medicine,
marketing, Classification of
Compute delta_wi bankruptcy consumer
for all weights prediction, spending pattern
from hidden layer insurance, the New product
to output layer ; list goes on. The analysis
backward pass following are
Compute delta_wi examples of Identification of
for all weights where neural customer
from input layer to networks have characteristics
hidden layer ; been used. Sale forecasts
backward pass Accounting
4. Predictin s examples of
REVIEW NEURA g commercial
OF L employe behavior
LITERAT NETW es applications for
URE ORK performa neural networks.
REPORTI PERFO nce and These include;
NG RMAN There Determining
Human CE: are personnel resource
resources numerou requirements
5. select a Neural networks can
ADVANT training of be implemented in
AGES OF data set interacti
NEURAL which ons
6. and between
DESIGN noisy factors. parallel hardware
PROBLE data.
MS: There is rarely one
Designin right tool to use in
fully g Neural data mining;
NETWOR describes Network
KS: the s using
6.
There are problem Genetic
When an element of
no general to be
the neural network
methods to solved. 4. fails, it
determine Independ Ease of it is a question as to
the ence
maintena what is available and
1. from
nce: what
High prior
Neural
Accuracy: assumpti networks
Neural ons:
can be
networks Neural Algorith can continue without
are able to ms any problem by their
optimal networks gives the best
number of do not results. Many articles,
neurones make a in addition
necessary priori updated
for assumpti with
ons fresh
data, parallel nature.
approximat 7. making to those mentioned in
e complex SOLUT them this paper, consider
non-linear IONS useful neural
mappings TO
Neuro-
solving anyIMPRO
Fuzzy
problem. VE
ANN Systems
Noise
Tolerance: networks to be a
Neural promising data
networks about the for mining tool.
are very distributi dynamic
flexible on of the environ
with data, or ments.
respect to the form 8.
41
incomplete PERFO CONCL
, missing RMAN USION:
It is CE:
difficult to 5.
Journal of Theoretical and Applied Information Technology

2005 - 2009 JATIT. All rights reserved.

www.jatit.org
Swami, A., Fundamentals
Artificial Neural Database of Neural
Networks offer Mining: A Networks:
qualitative Performance Architectures,
methods for Perspective, Algorithms and
business and IEEE Applications,
economic systems Transactions on Prentice-Hall,
that traditional Knowledge and New Jersey,
quantitative tools Data USA.
in statistics and Engineering,
econometrics pp. 914-925, Haykin, S.,
cannot quantify December 1993 Neural
due to the Networks,
complexity in Berry, J. A., Prentice Hall
translating the Lindoff, G., International
systems into Data Mining Inc., 1999
precise Techniques,
mathematical Wiley Khajanchi,
functions. Hence, Computer Amit, Artificial
the use of neural Publishing, Neural
networks indata 1997 (ISBN 0- Networks: The
mining is a 471-17980-9). next intelligence
promising field of
research especially Berson, Data Zurada J.M.,
given the ready Warehousing, An
availability of Data-Mining & introduction to
large mass of data OLAP, TMH artificial neural
sets and the networks
reported ability of Bhavani,Thura- systems, St.
neural networks to is-ingham, Paul: West
detect and Data-mining Publishing
assimilate Technologies,Te (1992)
relationships chniques tools
between a large & Trends,
numbers of CRC Press
variables. researchers use
them, in
Bradley, I., particular those
In most cases Introduction to
neural networks with statistical
Neural backgrounds.
perform as well or Networks,
better than the Thus, neural
Multinet networks are
traditional Systems Pty Ltd
statistical becoming very
1997. popular with
techniques to
which they are data mining
compared. Fayyad, Usama, practitioners,
Resistance to Ramakrishna particularly in
using these black Evolving Data medical
boxes is mining into research,
gradually solutions for finance and
diminishing as Insights, marketing. This
more communications is because they
of the ACM 45, have proven
no. 8 their predictive
REFERENCES power through
Fausett, comparison
Agrawal, R., Laurene (1994), with other
Imielinski, T.,
statistical they are widely models with
techniques using accepted in user-friendly
real data sets. Due industry. As interfaces the
to design problems software attraction to
neural systems companies neural networks
need further develop more will continue to
research before sophisticated grow.

42

Vous aimerez peut-être aussi