Classifier Model using Artificial Neural Network

International Journal of Engineering, Business and Management (IJEBM)
ISSN: 2456-7817
[Vol-7, Issue-4, Jul-Aug, 2023]
Issue DOI: https://dx.doi.org/10.22161/ijebm.7.4
Article DOI: https://dx.doi.org/10.22161/ijebm.7.4.11
Int. j. eng. bus. manag.
www.aipublications.com Page | 69
Classifier Model using Artificial Neural Network
Inderjit Kaur1
, Dr. Pardeep Saini2
1
Research Scholar, Sunrise University, Alwar, Rajasthan, India
2
Professor, Sunrise University, Alwar, Rajasthan, India
Received: 12 Jul 2023; Received in revised form: 09 Aug 2023; Accepted: 16 Aug 2023; Available online: 25 Aug 2023
©2023 The Author(s). Published by AI Publications. This is an open access article under the CC BY license
(https://creativecommons.org/licenses/by/4.0/)
Abstract— When it comes to AI and ML, precision in categorization is of the utmost importance. In this
research, the use of supervised instance selection (SIS) to improve the performance of artificial neural
networks (ANNs) in classification is investigated. The goal of SIS is to enhance the accuracy of future
classification tasks by identifying and selecting a subset of examples from the original dataset. The purpose
of this research is to provide light on how useful SIS is as a preprocessing tool for artificial neural network-
based classification. The work aims to improve the input dataset to ANNs by using SIS, which may help with
problems caused by noisy or redundant data. The ultimate goal is to improve ANNs' ability to identify data
points properly across a wide range of application areas.
Keywords— Artificial Neural Network, supervised instance selection, Data classification, machine
learning.
I. INTRODUCTION
The primary goal of any data classifier is to appropriately
categorize patterns into one of many groups that may or may
not be known. The field of data classification has attracted
neural networks because to its impressive not-linear
function approximation and adaptive learning capabilities.
The first step in any data classification process is to create
a model that stands in for the various data classes, and the
second is to use a model that was specifically made for
classification.
These fundamentals of artificial neural networks
demonstrate the sufficiency of a Feed Forward Neural
Network in tackling difficult data classification problems.
The development of classification models using ANNs is
similarly fraught with difficulties.
Training samples for the k-Nearest Neighbor Data
Classification technique are stored uniformly across n
dimensions. When an unknown sample is provided, the
algorithm calculates the Euclidean distance between the
sample and the unknown and then searches the pattern space
for the k samples that are closest to the unknown.
Classification schemes that use neighboring nations as
examples retain all training samples and wait to create a
classification until a new sample is classified. When
comparing an unlabeled sample to a large pool of potential
neighbors, they may rack up hefty computational costs.
II. LITERATURE REVIEW
Narender Kumar, (2020) In machine learning, you may go
one of two ways: supervised or unsupervised. Supervised
learning may be used to the classification approach. Among
the many classification methods available, the Artificial
Neural Network stands out as the most widely used. Neural
networks are useful for classifying data and creating
models, but their accuracy is debatable. The artificial neural
network is optimized to provide more precise and timely
results. The Bat Algorithm is a metaheuristic algorithm that
may be used with ANN to create a hybrid system.
Optimizing the neural network has several benefits,
including better classification accuracy, better data
interpretation, lower costs, less time spent, etc. In this
research, we evaluate the ANN Back propagation model's
results for medical diagnosis against those of our proposed
ANN-Bat model. Results showed that the ANN-Bat
approach was superior, cutting delivery times and
improving precision.
Wanto.et.al Anjar (2017) The creation of artificial neural
networks is a computing paradigm that borrows heavily
from the biologically inspired structure of intelligent brains.

Kaur and Saini/ Classifier Model using Artificial Neural Network
There are many uses for artificial neural networks in the
computing world. One of them stores information used for
making predictions. Since the back spread algorithm can
learn from historical data and identify data patterns,
artificial neural networks of the back spread kind are quite
popular. It's possible to analyze and forecast future events
based on this background reproduction pattern. The Human
Development Index from 2011-2015 is the source of
information for this analysis. North Sumatra statistics from
the Central Bureau of Statistics. The research used the 3-8-
1, 3-18-1, 3-28-1, 3-16-1, and 3-48-1 architectural models.
With an epoch of 5480 iterations and 0.0006386600 with
error 0.001 to 0.05, Model 3-48-1 in architectural design has
the highest accuracy of the five models, at 100%. Therefore,
when employed for data prediction, the 3-48-1 back
propagation approach is adequate.
MARCIN BLACHNIK (2019) Preprocessing techniques
such as selecting instances and characteristics may
drastically decrease computational complexity and improve
prediction accuracy. Despite the widespread academic
interest in finale prediction models, only few authors have
delved into set selection methods. To fill this need, this
research looks at four sets specifically designed for instance
selection: bagging, function bagging, adaboost, and extra
noise. This is the first time that last one has been seen in
print. The study relies on an empirical comparison of 43
datasets and 9 fundamental instance selection procedures.
There are three different types of testing. In the first, the
impact of ensembles on the compression relation is shown
using a single dataset for evaluation. The second case is
concerned with optimizing for predicted accuracy, whereas
the third case involves balancing many criteria, including
data compression. The gathered data demonstrates that,
with the exception of unstable methods like CNN and IB3,
instance selection ensembles improve upon the fundamental
instance selection algorithms, although with a compression
cost. In most cases, Bagging and AdaBoost are superior.
Specifically, 1NN, kNN, and SVM are tested and compared
in the studies. We also discover that the prediction accuracy
of robust classifiers (kNN and SVMs) based on data filtered
by installation (including ensembles) decreases when
compared to the results obtained in the whole training set
for the training of these classifiers.
Sonam Saxena.et.al (2019) In recent years, data mining has
seen rapid growth and widespread use of associated
technologies. Quick conclusions may be reached by using it
to analyze past data. A formalized method of making
decisions has the potential to improve data protection as
well. An example data mining application is shown in the
material on offer. The proposed use of data mining enhances
data protection. As a result, we consider the problem of how
to classify URLs. In this research, we propose using
association rule-mining technology to resolve URL
classification, however the supervised learning technique
might also be useful. Phishing and legitimate websites'
URLs may be analyzed using this technique. It is proposed
to apply a classification strategy based on rules to this
domain. This approach may be used to classify URL
information based on calculated association criteria. The
inspiration for this originates from the usage of apriori
algorithms for the creation and categorization of phishing
URLs. The computational and memory requirements of the
apriori method for generating candidate sets are high. We
use the FP-Tree method, which efficiently generates
lightweight association rules. This method has potential use
in the development of phishing toolbars. This approach is
used to compare the results of the Phish tank dataset to those
of other datasets. The results indicate that the suggested
approach requires less mental effort and storage space.
There will soon be a more efficient and less cumbersome
approach for classifying potential phishing URLs.
Jonathan Schmidt (2019) Among the many fascinating
new techniques in materials science, machine learning
stands out. It has been shown that basic and applied research
may benefit considerably from this suite of statistical
techniques. Recent years have seen a proliferation of
research into using machine learning to semiconductors. We
review and discuss the most recent studies on the topic. We
introduce the fundamentals of machine learning, including
algorithms, descriptors, and databases for materials science.
We continue to detail other machine learning-based
strategies for locating stable materials and predicting their
crystal structure. We give studies on several strategies for
replacing fundamental principles with machine learning, as
well as many quantitative linkages between structures and
characteristics. We investigate the potential of active
learning and surgical optimization to improve rational
design and associated processes. Two perennial issues with
machine learning models are their lack of interpretability
and physical understanding. As a result, we discuss the
significance of interpretability in materials science and the
different facets of this concept. Finally, we provide
solutions to a variety of computational materials science
problems and suggest directions for further study.
III. ARCHITECTURE OF FEED FORWARD
NEURAL NETWORK
An artificial neural network is a paradigm for processing
data that takes cues from the brain. It's made up of a network
of neurons all working together to find a solution to a certain
problem. The architecture of a three-layer feed forward
neural network (FFNN) is shown in Figure 1.

All of the designs in this subclass of neural networks have
one thing in common: they all use unidirectional
connections between neurons in successive layers. That is
to say, information may go in just one direction (the
"forward direction") via a given set of branches and links.
The weights of the connected branches may be adjusted
according to a user-defined learning policy. Neuronal
connections to other architectural layers are not made
possible by feedback networks. The neuron's response is
generated by feeding the linear combiner's output (the
neuron's activity level) into a non-linear active function f (.).
Fig.1: Architecture of Feed Forward Neural Network
The network's neuronal activity typically falls between -1
and 1, however the range [0, 1] is useful in certain contexts.
There are really three distinct layers in Figure 1; The signal
for the "second" layer neurons is input in the "third" layer
(or the output layer), and no computations are done in the
"first" layer. The network's responsiveness is measured by
what comes out of the last layer, the output layer.
Non-linear mapping between inputs and outputs is possible
in this network. While there may be many theoretically
possible hidden levels in architecture, in practice just one or
two are often used. To approximate non-linear mapping, all
that's needed is a multi-layer perceptron with a single
hidden layer and enough neurons. The identification of a
large enough number of neurons to achieve the required
approximation accuracy is notoriously challenging in
practice. Therefore, the trial-and-error method is used to
determine the density of the hidden layer.
IV. LEARNING IN NEURAL NETWORK
The ability to adjust to new conditions is the primary source
of their resilience and strength. Throughout the process of
readjustment, they construct mental models using
information about their surrounding environment. These
mental representations are written down as various
"structured" vectors of importance. Learning algorithms
describe a process that is architecture-dependent and entails
the encoding of data input into weights to generate these
internal models. Strengthening and weakening connections
is how learning occurs.
Postsynaptic channels in biological learning systems are
affected by the efficiency of the synapse, both in terms of
the amount of neurotransmitters produced by a synaptic
terminal and the physical shape of the axon-dender junction.
In artificial systems, learning alters the model's synaptic
weights.
Data is the primary engine of most learning. Input-output
batteries representing data from a (perhaps unknown)
probability distribution are possible. In this scenario, the
output pattern may reveal the system's reaction to a given
input pattern, and the learning task would then be to
approximatively determine the unknown function. It's also
possible that learning will be difficult since the data
contains patterns that naturally cluster into several unknown
classes.
For the purpose of training and testing neural networks,
several different learning algorithms are at your disposal. In
this research, a backpropagation-based learning algorithm
is developed for use in training and evaluating the neural
feed network. Details of the back propagation algorithm are
laid down below.
V. BACK PROPAGATION ALGORITHM
The Back Propagation learning method is a step-down
strategy for minimizing the mean square error between the
observed and desired output of a multi-layer perceptron.
When training a network using back propagation, a non-
linear relationship is created between the input and output
values. To account for the nonlinear relationship between
the input and output pairs, the network may adjust its
weights using the rear propagation strategy.
The method for background propagation includes:
Step 1. Weigh and offset initialization
Weights and node offsets are first set to arbitrary small
values.
Step 2. Present vector input and output desired
Put forth the input x as a continuous vector and specify the
output you seek. d. All members of the vector output are 0
unless they belong to the current input class.
Step 3. Compute current outputs
Get the vector of output values right now, then apply the
sigmoid nonlinearity to them.
𝑓(𝑛𝑒𝑡𝑖) =
1
1 − 𝑒−𝑛𝑒𝑡
Step 4 Adapt weights Adjust weights by 𝑤𝑖𝑗(𝑡 + 1) =
𝑤𝑖𝑗(𝑡) + 𝑛𝑗
&
𝑥𝑖

Where the output of the node I, η is the sensitivity of node
j, and the learning rate constant. If node j is a destination
node, then 𝛿𝑗 = 𝑓(𝑛𝑒𝑡𝑗)(𝑑𝑗 − 𝑦𝑗)
Where 𝑓(𝑛𝑒𝑡𝑗) netj's estimated activation function is the
target value for node j's output, whereas yj represents the
actual value. Sensitivity is defined as where j is the index of
the node and if it is an internal node.
(3.4)
where k is the accumulated weight of all nodes above layer
j. Using the LMS training criteria function and the chain
derivation procedure, we can derive the update equations.
Step 5. Repeat by going to step 2
If the shift in the exercise criterion is smaller than a set
threshold, the workout might be considered complete.
When the training error in one validation set is small
enough, the cross-validation approach ceases.
After being trained, networks with fixed weights may be
able to provide an output for a given input. Once the
network has been trained, it may be used as a classifier
model in any engineering context.
VI. MUTUAL INFORMATION-BASED
FEATURE SELECTION
Concept Of Mutual Information
Entropy is a measure of the average amount of uncertainty
around a random experiment. Let Y be a discrete random
variable with potential values yi, I = 1, 2,... NY, and let
Prob(Y=yi) = Pi be its probability distribution function to
characterize a random experiment. Then, the formula gives
a definition of the entropy of the random experiment,
𝐻(𝑌) = − ∑ 𝑃𝑖𝑙𝑜𝑔𝑝𝑖
𝑁𝑦
𝑖=𝑗
The initial entropy of a random experiment may be
decreased if we know more information X about it. If you
know X, then the conditional entropy of a random
experiment is
𝐻(𝑌|𝑋) = ∑ 𝑃𝑗
𝑁𝑦
𝑗−1
(∑ 𝑝(𝑦𝑖¿ 𝑥𝑗) log 𝑃 (𝑦𝑖¿ 𝑥𝑗)
𝑁𝑦
𝑗−1
)
Where Pj is the probability distribution function of X with
possible values xj, j=1,2,…,Nx and P(yi/xj) is the likelihood
that yi will occur if xj does. The conditional entropy is
always less than or equal to the original entropy. For any
two sets of information Y and X, the mutual information
I(Y; X) is the amount by which the entropy (uncertainty) is
reduced:
I (Y; X) = H(Y) – H(Y/X)
As a result, the mutual information level lowers the typical
degree of uncertainty regarding the experiment's random
outcome Y. Mutual information is the symmetrical metric.
In other words, the amount of knowledge gained about Y
after seeing X is the same as the amount of knowledge
gained about X after seeing Y. X is the raw data, and Y is
the final class label for this function selection problem.
Computation Of Mutual Information
To compute mutual information, we must use the best
information at our disposal, which is the histogram of data,
to represent the probability distribution of variables that
does not exist in reality. Here are the steps required to derive
the inverse data from the training data histogram:
Step 1: Sort the output patterns from most numerous to least
many, and then divide the sorted patterns into equality in Ny
classes.
Step 2: If you don't know anything about the input variable,
you can figure out the initial entropy of the output Y.
Step 3: Separate X1 into Nx equal subsets based on
descending pattern similarity.
Step 4: Determine Y's entropy if and only if we know the
value of X1.
Step 5: Find out what knowledge about X1 Y has that X1
does not have.
Step 6: To account for the remaining variables, repeat Steps
3–5.
VII. CONCLUSION
A classifier with strong generalizability might be
constructed using a neural network with optimal topology.
It is possible that the cutting process will reveal the optimal
structure of neural networks. Act swiftly to find a solution
by starting with a large network and gradually shrinking it
to a smaller network with the goal of increasing generality.
When pre-processing and/or pruning improve the
classifier's performance, the data used to train it may be as
basic as a set of rules for making a classification. The
classification rules may be extracted with the aid of the rule
extraction technique from the cut network, which is easier
to comprehend as a condensed trained network. The thesis
focuses on essential principles that facilitate the efficient
use of neural networks in the creation of the classifier. It has
led to advancements in discretization methods, pattern
recognition, and neural network design. This discrete
algorithm's findings show that the proposed discrete system
requires less discrete time and produces more accurate
classifications with a less number of intervals.

REFERENCES
[1] Anjar Wanto.et.al 2017,” Analysis of Artificial Neural
Network Accuracy Using Backpropagation Algorithm in
Predicting Process (Forecasting),” International Journal Of
Information System & Technology Vol. 1, No. 1, (2017), pp.
34-42
[2] Narender Kumar,2020,” Classification using Artificial
Neural Network Optimized with Bat Algorithm,”
International Journal of Innovative Technology and
Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-
9 Issue-3, January 2020
[3] Marcin Blachnik 2019,” Ensembles Of Instance Selection
Methods: A Comparative Study,” Int. J. Appl. Math. Comput.
Sci., 2019, Vol. 29, No. 1, 151–168 DOI: 10.2478/amcs-
2019-0012
[4] Sonam Saxena.et.al 2019,” A Data Mining approach to Deal
with Phishing URL Classification Problem,” International
Journal of Computer Applications (0975 – 8887) Volume 178
– No. 41, August 2019
[5] Jonathan Schmidt 2019,” Recent advances and applications
of machine learning in solid-state materials science,” npj
Computational Materials volume 5, Article number: 83
(2019)
[6] R. Vijaya Kumar Reddy.et.al 2018,” A Review on
Classification Techniques in Machine
Learning,”InternationalJonroual of Advance Research in
Sciences and Engineering Volumn no.7
[7] Anu Sharma.et.al. 2017,” Literature Review and Challenges
of Data Mining Techniques for Social Network Analysis,”
Advances in Computational Sciences and Technology ISSN
0973-6107 Volume 10, Number 5 (2017) pp. 1337-1354
[8] statistical learning theory; optimisation theory; financial
econometrics; support vector machine; SVM; kernel
methods. DOI: 10.1504/IJBIDM.2019.10019195
[9] Dharmender Kumar 2017,” Classification Using ANN: A
Review,” International Journal of Computational Intelligence
Research ISSN 0973-1873 Volume 13, Number 7 (2017), pp.
1811-1820
[10] Tameru Hailesilassie 2016,” Rule Extraction Algorithm for
Deep Neural
[11] Networks: A Review,” (IJCSIS) International Journal of
Computer Science and Information Security, Vol. 14, No. 7,
July 2016
[12] María Pérez-Ortiz.et.al. 2016,” A Review of Classification
Problems and Algorithms in Renewable Energy
Applications,” Energies 2016, 9, 607;
doi:10.3390/en9080607 www.mdpi.com/journal/energies
[13] Alvaro Osornio-Vargas.et.al. 2017,” A systematic review of
data mining and machine learning for air pollution
epidemiology,” BMC Public Health volume 17, Article
number: 907 (2017)
[14] Abhishek, K., M. P. Singh, S. Ghosh and A. Anand (2012).
Weather forecasting model using artificial neural network,
Procedia Technology, Vol. 04, pp. 311-318.
[15] Ali-Khashashneh, E. A. and Q. A. Al-Radaideh (2013).
Evaluation of discernibility matrix based reduct computation
techniques, 5 th International Conference on Computer
Science and Information Technology - IEEE, Amman, pp.
76-81, Jordan.

Classifier Model using Artificial Neural Network

More Related Content

Similar to Classifier Model using Artificial Neural Network (20)

More from AI Publications (20)

Recently uploaded (20)

Classifier Model using Artificial Neural Network