IRJET-Performance Enhancement in Machine Learning System using Hybrid Bee Colony based Neural Network

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 01 | Jan-2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 6.171 | ISO 9001:2008 Certified Journal | Page 931
Performance Enhancement in Machine Learning System using Hybrid
Bee Colony based Neural Network
S. Karthick 1*
1 Team Manager, Sea Sense Softwares (P) Ltd., Marthandam, Tamil Nadu, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Metric learning and data prediction is a notable
issue in many data mining and machine learningapplications.
In recent advances many researches have been proposed for
the data prediction. Still there is a problem such as reduced
prediction accuracy, computational timecostetc.Henceinthis
paper a novel Hybrid Bee Colony based Neural Network
(HBCNN) is proposed for the data prediction in data mining.
The proposed technique is segmented into three phases, in the
first phase the raw input data is pre-processed to make it
suitable for computation. The second phase is training of the
proposed HBCNN technique is which the neural network is
trained using bee colony algorithm. The last phase is dynamic
testing in which the neural network is undergonerapidtesting
with dynamic input for the validation and testing of trained
ANN. The proposed technique is tested using cancer dataset
and the performance is compared with KNN technique.
Key Words: Metric Learning, neural network, bee
colony, data prediction
1.INTRODUCTION
Artificial Neural Network (ANN) is often a new industry
connected with computational science which integrates the
various strategies for difficulty resolution which cannot be
consequently effortlessly described without the help of an
algorithmic conventional concentration.TheparticularANNs
stand for big as well as diverse classes connected with
computational types [1]. Together with biological template
modules these people are created basically by more or less
comprehensive examples. The particular general
approximates as well ascomputational typeshavingspecific
traitssuch as the ability to learn or even adapt, to prepareor
to generalize data are recognized as the particular Artificial
Neural Networks. Contemporary development of a (near)
optimal network architecture is carried out by means of
people skilled together with use of a monotonous learning
from trial and error process. Neural networks are generally
algorithms regarding optimization aswellasfindingout,and
are dependent freely with principles prompted merely by
researching on the characteristics from the brain. For
optimization as well as finding out neural networks in
tandem with genetic algorithms, there are generally two
techniques, each using its own advantages coupled with
weaknesses. By means of different paths, the two processes
are commonly evolved [2, 3].
Optimization algorithms, in addition tobeingcalledLearning
algorithms, are, according to a number of features of
scientific advancement, usuallyknownasGeneticalgorithms.
An easy method involving encoding answers to the issue in
chromosomes, an assessment purpose which on return
starts to attain Chromosome directed at the idea, involving
initializing population of chromosomes, workers that could
be put on parents whenever they reproduce to correct the
genetic composition are the five expected criteria.
Integration could be of mutation, crossover in addition to
site-specific workers. Parameter configurations for the
particular criterion are the workers or can be anything else
[4]. The criteria can develop populations involving much
better ones in addition to the much better individuals,
converging eventually in effect close to an international
ideal, every time a genetic algorithm is actually run having a
portrayal which usefully encodes a solution to a problem in
addition to workers that can generate much better young
onesthrough good parents[5]. In numeroussituations,with
regard to performing the particular optimization on the
standard workers, mutation in addition to crossover, is
usually adequate [6]. Genetic algorithms can assist like a
black box function optimizer certainly not requesting any
kind of know-how about computersof the particulardomain
in these cases. However, understanding of the particular
domain is frequently exploited to raise the particulargenetic
algorithms performance with the incorporation involving
new workers highlighted in this particular paper [7].
In neural network, functionality enhancement partition
space can be a space which is used to classify data sample
right after the test is actually mapped by neural network[8].
Centroid can be a space with partition space and also
denotes the particular middle of the class. In traditional
neural network functionality, location of centroids and the
partnership involving centroids and also classes are
generally fixed personally [9]. Furthermore, with reference
to the quantity of classes variety of centroids is actually set.
To locate optimum neural network this particular set
centroid restriction minimizes the risk [9]. Genetic
algorithms (GAs) have emerged to be practical software for
the heuristic alternative of sophisticated discrete
optimization issues. Particularly, there has been large
fascination with their use in the most effectivearrangingand
timetabling issues[10]. However, nowadaystherehavebeen
several attempts made to become listed on both systems.
Neural networks could be rewritten since a type of genetic
algorithm is termed as some sort of classifier program and
also vice-versa [11]. Genetic algorithm is meant for training
feed forward networks. It doesnot merely work withitstask
but it does perform rear distribution, the normal training

criteria [12, 13]. This accomplishment comes from tailoring
the particular genetic algorithm to the domain of training
neural networks.
Although the ANN features enjoy several advantages it has
got several complications similar to convergence inaddition
to receiving throughout neighborhood minima during
training by means of back propagation [14]. That is why the
analysts are engaged to produce a new finest protocol to
train the ANN. On this routine optimization protocol,
algorithms similar to PSO, CS etc have been utilized to
enhance the overall performance however it certainly does
not fulfill the overall performance prerequisites. On this
impression our system is also thought out to be able to get
over a real issue during training involving ANN. Thus, we
proudly propose the novel AGCS technique for training the
ANN.
2. PROPOSED HBCNN TECHNIQUE FOR MACHINE
LEARNING SYSTEM
Metric learning and Data prediction is a notable issue in
many data mining and machine learning applications. In
recent advancesmanyresearcheshavebeenproposedforthe
data prediction. Still there is a problem such as reduced
prediction accuracy, computational time cost etc. Hence a
novel hybrid HBCNN is proposed for the data prediction in
data mining.The architecture of the proposedsystemisgiven
in fig 1.
1
2
3
4
5
6
7
8
CancerDataAttribute
Database Pre-processing Feature extraction Classification Output
Disease
Fig -1: Architecture of the proposed system
The proposed system includes four phases they are;
1. Data Collection
2. Data Preprocessing
3. ANN-ABC Predictor Modelling
4. Dynamic Testing
2.1 Artificial Neural Network
An ANN is a programmed computationalmodelthat aims
to duplicate the neural structure and the functioning of the
human brain. It is made up of an interconnected structure of
artificially produced neurons that function as pathways for
data transfer. Artificial neural networks are flexible and
adaptive, learning and adjusting with each different internal
or external stimulus. Artificial neural networks are used in
sequence and pattern recognition systems, data processing,
robotics and modeling. The ANN consists of a single input
layer, and a single output layer in addition to one or more
hidden layers. All nodesare composed of neurons except the
input layer. The number of nodes in each layer varies
depending on the problem. The complexity of the
architecture of the networkis dependentuponthenumberof
hidden layers and nodes. Training an ANN is to find a set of
weights that would give desired values at the output when
presented with different patterns at its input. The two main
process of an ANN is training and testing.
The total of eight attributes or input is given as the input
for the ANN and the corresponding class or disease can be
obtained as the output of the ANN. Thus the proposed ANN
architecture contains eight inputs (eight attributes) and
corresponding disease output. The main two process of a
classification algorithm is training and testing.
In training the input as well as the output will be defined
and the appropriate weight is fixed so that the classifier
(ANN) can able to predict the apt object (disease) in the
testing phase. Hence the training phase is the major part in a
common classifier algorithm. In the proposed system the
artificial bee colony (ABC) algorithm is used for the training.
The process involved in the training is given as follows and
the architecture of the back propagation neural network
(BPNN) is given in the fig 2.

F1
F2
F3
F4
F5
C
F6
F8
F7
wI
11
wI
12
wI
13
wI
14wI
15
wI
21
wI
22
wI
23
wI
24
wI
25
wI
31
wI
32
wI
33
wI
34
wI
35
wI
41
wI
42
wI
43
wI
44
wI
45
wI
51 wI
52
wI
53
wI
54
wI
55
wI
61
wI
62
wI
63
wI
64
wI
65
wI
71 wI
72
wI
73
wI
74
wI
75
wI
81
wI
82
wI
83
wI
84
wI
85
wO
1
wO
2
wO
3
wO
4
wO
5
Fig -2: Proposed back propagation neural network
Theproposed ANN consistof eightinputunits,oneoutput
units, and M hidden units (M=5). First, the input data is
transmitted to thehidden layer and then, to theoutputlayer.
This is called the forward pass of the back propagation
algorithm. Each node in the hidden layer gets input from the
input layer, which are multiplexed with appropriate weights
and summed. The output of theneuralnetworkisobtainedby
the Eqn. (1) given below.





M
j
N
i
I
iji
O
j
wF
w
C
1
1
)exp(1
(1)
In Eqn. (1), ‘ iF ’ is the
th
i input value and ‘
O
jw ’ is the
weights assigned between hidden and output layer, ‘
I
ijw ’ is
the weight assigned between input andhiddenlayerandMis
the number ofhidden neurons.Theoutputofthehiddennode
is the non-linear transformation of the resulting sum. Same
process is followed in the output layer. The output values
from the output layer are compared with target values and
the learning error rate for the neural network is calculated,
which is given in Eqn. (2).
 2
2
1
CYfit k  (2)
In Eqn. (2), k is the
th
k learning error of the ANN, Y is
the desired output and C is the actual output. The error
between the nodes is transmitted back towards the hidden
layer. This is called the backward pass of the back
propagation algorithm. Then the training is repeated for
some other training dataset by changing the weights of the
neural network. The learning error is considered as the
fitness function in ABC for the error minimization.
B. Artificial Bee Colony Algorithm
Artificial Bee Colony (ABC) algorithm is a comparatively
new technique proposed by Karaboga. ABC is encouraged by
the foraging behaviour of honey bee swarms. In the ABC
algorithm, the colony consists of three kinds of bees namely:
 Employee bee
 Onlooker bee
 Scout bee
Among these three kinds of bees, the employee bee and
the onlooker bee are the employed bees, whereas, the scout
bee is an unemployed bee.The numberof employeebeesand
onlooker beesare said to be equal in ABCalgorithm.Thefood
sources are considered as the possible solutions for a given
problem and the nectar amount of the food source is the
relative fitness of that particular solution. The population of
the colony is twice the size of the food sources. The number
of food sources represents the position of the possible
solutions of the optimization problem and thenectaramount
of a food source represents the quality (fitness) of the
associated solution.The pseudo code of the ABCalgorithm is
given below.

Initialize food sources
Repeat
Employed Bee Phase:
Send the employed bees to the food sources assigned
to them.
Determine the amountof nectar (fitnessvalue)inthe
food source.
Calculate the probability value of the food sources.
Onlooker Bee Phase:
Select the food sources discovered by the employee
bee based on the probability value.
Find the neighboring food source and estimate its
nectar amount.
Compare both the food sources and select the food
source with better fitness.
Scout Bee Phase:
Select the food sources randomly and replace the
abandoned food source with the new food source.
Memorize the best food source.
Until (Requirements are met)
In the initial step of ABC algorithm, a set of food sources
are selected randomly by the employed bee and their
corresponding nectar amounts (i.e. fitness value) are
computed.The employee bee sharesthese detailsto another
set of employed bees called as onlooker bees. After sharing
this information, the employee bee visits the same food
source again and then, finds a new food source nearby using
the Eqn. (3).
)( ,,,,, jkjijijiji xxxv  (3)
Where, k and j are random selected index that represent
the particular solution from the population.  is a random
number between [-1,1].
When new neighbouringfoodsourcesaregenerated,their
fitness values are calculated and the employee bee applies
greedy selectiontomake a decisiononwhethertoreplacethe
existing food sourceinmemoryusingnewfoodsourceornot.
Then, the probability value Pi is calculated for each food
source based on its fitness amount using the following Eqn.
(4).

 2/
1
CS
i
i
i
i
fit
fit
P (4)
Where, CS is the colony size.
3. Results and Discussion
The proposed system is tested using cancer dataset and
the results are compared with conventional techniques. The
ALL / AML dataset for this experimental analysis is collected
from online. After the normalization the randomly chosen
sampleis dividedinto three categoriessuch astraining,cross
validation and testing data sets. The training data set is used
forlearning the network. Crossvalidation is usedto measure
the training performance during the training as well as to
stopthe training if necessary.TheLeukaemiacancercontains
four types. In our system we consider only two major types
namely ALL (Acute Lymphoblastic Leukaemia) and AML
(Acute Myeloid Leukaemia). For this training purpose we
consider 38 patients who are divided into two clusters with
25 and 13 patientsfor ALL and AML respectivelywithoverall
7129 genes.
The performance of the proposed system is compared
base on the execution time and accuracy of the prediction.
Execution time is the time in which a single instruction is
executed. It makes up the last half of the instruction cycle.
Accuracy is used todescribe the closenessof a measurement
to the true value. When the term is applied to sets of
measurements of the same measured, it involves a
component of random error and a component of systematic
error. In this case trueness is the closeness of the mean of a
set of measurement results to the actual (true) value and
precision is the closenessof agreementamongasetofresults.
Table 1: performance analysis table for k=1
K=1
Techniques
Proposed Existing System Cosine Based System
Data
Set
accuracy
execution
time
accuracy
execution
time
accuracy
execution
time
iris 96.00%
0.148
seconds.
96.00%
0.16
seconds.
94.67%
0.022
seconds
glass 71.03%
0.069
seconds.
67.29%
0.11
seconds.
67.29%
0.023
seconds.
vowel 96.77%
0.174
seconds.
96.36%
0.51
seconds.
91.72%
0.119
seconds.
cancer 67.14%
0.124
seconds.
63.59%
0.88
seconds.
65.48%
0.116
seconds.
letter 93.95%
3.569
seconds.
94.55%
4.878
seconds.
94.08%
9.692
seconds.
DNA 73.38%
1.886
seconds.
73.38%
1.938
seconds.
71.25%
3.013
seconds.
For k=1 performance analysis is compared with the
existing technique. For the cosine based system execution
time is evaluated with the proposed and the existing
strategies. Comparison analysis is evaluated for the existing
and the cosine based system. This comparison is carried out
for the iris, glass, vowel, cancer, letter and DNA.
strategies.

Fig -3: Performance analysis chart for k=2
Figure 3 shows the performance analysis of the chart for
k=2.Comparison analysisis evaluated for the existingandthe
cosine based system. This comparison is carried out for the
iris, glass, vowel, cancer, letter and DNA.
k=1 forthe execution time. Comparison analysis isevaluated
forthe existing andthe cosine basedsystem.Thiscomparison
is carriedout for the iris, glass, vowel, cancer, letter andDNA.
strategies.
k=3. Comparison analysis is evaluatedfortheexistingandthe
cosine based system. This comparison is carried out for the
iris, glass, vowel, cancer, letter and DNA.
k=1 forthe execution time. Comparison analysis isevaluated
forthe existing andthe cosine basedsystem.Thiscomparison
is carriedout for the iris, glass, vowel, cancer, letter andDNA.
Fig -4: Performance analysis chart the execution time for K=2

Fig -5: Performance analysis chart for k=3
Fig -6: Performance analysis chart the execution time for K=2
4. Conclusion
Machine learning is the motivated research topic in the
recent decade. Metric learning become a challenging task in
machine learning. Hence a novel procedure for the metric
learning is proposed in thispaper.The proposedsystemuses
two stages of execution, in the first stage semi supervised
clustering and second stage prediction is performed. The
hierarchy forest clustering technique is utilized for the semi
supervised clustering and KNN technique for the prediction.
The performance of the system is verified based on the
accuracy and execution time. The performance analysis
outperforms the existing techniques and proves its
effectiveness for the metric learning in machine learning.
REFERENCES
[1] Mantas Lukosevicius, and Herbert Jaeger, “Reservoir
Computing Approaches to Recurrent Neural Network
Training,” Computer Science Review, Vol. 3, 2009, pp.
127–149.
[2] David E. Moriarty, Alan C. Schultz, and John J.
Grefenstette, “Evolutionary Algorithms for
Reinforcement Learning,” JournalofArterialIntelligence
Research, Vol. 11, 1999, pp. 241-276.
[3] Lucas Negri, AdemirNied, Hypolito Kalinowski, and
Alexander Paterno, “Benchmark for Peak Detection
Algorithms in Fibre Bragg Grating Interrogation and a
New Neural Network for itsPerformanceImprovement,”
Sensors, Vol. 11, 2011, pp. 3466-3482,
[4] David J. Montana and Lawrence Davis, “Training Feed
forward Neural NetworksUsing Genetic Algorithms,”In
Proceedings of the 11th international joint conference
on Artificial intelligence, Vol. 1, 1989, pp. 762-767,
[5] William B. Langdon, Robert I. McKay and Lee Spector,
“Genetic Programming,” International Series in

Operations Research & Management Science, Vol. 146,
2010, pp. 185-225,
[6] Swagatam Das, Ajith Abraham, UdayK.Chakraborty,and
Amit Konar, “Differential Evolution Using a
Neighborhood-Based Mutation Operator,” IEEE
Transactions on Evolutionary Computation, Vol. 13,
2009, pp. 526-553,
[7] Adam M. Smith and Michael Mateas, “Answer Set
Programming for Procedural Content Generation: A
Design Space Approach,” IEEE Transactions on
Computational Intelligence and AI in Games, Vol. 3,
2011, pp. 187 -200,
[8] Lin Wang, Bo Yang, Yuehui Chen, Ajith Abraham,
Hongwei Sun, Zhenxiang Chen and Haiyang Wang,
“Improvement of neural networkclassifierusingfloating
centroids,” Vol. 31, 2012, pp. 433-454,
[9] [9] Lei Zhang, Lin Wang, Xujiewen Wang, Keke Liu, and
Ajith Abraham, “Research of Neural Network Classifier
Based on FCM and PSO for Breast Cancer Classification,”
Hybrid Artificial Intelligent Systems, Vol. 7208, 2012,
pp. 647-654,
[10] Tanomaru J., “Staff Scheduling by a Genetic Algorithm
with Heuristic Operators,” In Proceedings of the IEEE
Conference on Evolutionary Computation, 1995, pp.
456-461.
[11] Patrycja Vasilyev Missiuro, “Predicting Genetic
Interactions in Caenorhabditis Elegans using Machine
Learning,” MassachusettsInstitute of Technology,2010,
pp. 191-204
[12] Piotr Mirowski, Deepak Madhavan, Yann LeCun, and
Ruben Kuzniecky, “Classification of patterns of EEG
synchronization for seizure prediction,” Clinical
Neurophysiology, Vol. 120, 2009, pp. 1927–1940
[13] XS Yang and S. Deb, “Cuckoo search via Lévy flights,” In
Proceedings of World Congresson Nature &Biologically
Inspired Computing, 2009, pp. 210-214
[14] L. Mingguang and L. Gaoyang, “ArtificialNeuralNetwork
Co-optimization Algorithm based on Differential
Evolution,” In Proceedings of Second International
Symposium on Computational Intelligence and Design,
2009

IRJET-Performance Enhancement in Machine Learning System using Hybrid Bee Colony based Neural Network

More Related Content

What's hot (17)

Similar to IRJET-Performance Enhancement in Machine Learning System using Hybrid Bee Colony based Neural Network (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET-Performance Enhancement in Machine Learning System using Hybrid Bee Colony based Neural Network