Fake Currency Notes Detection using Supervised

2024 International Conference on Advances in Computing, Communication and Applied Informatics (ACCAI) | 979-8-3503-8944-9/24/$31.00 ©2024 IEEE | DOI: 10.1109/ACCAI61061.2024.10602135

Learning Methods
K. Selvakumar1, Premlatha K.R2, M. Tamil Thendral3, Sai Ramesh L4
1
Department of Computer Applications National Institute of Technology, Tiruchirappalli, India
2
Assistant Professor, Department of Information Technology, Thiagarajar College of Engineering, Madurai - 625 015.
3,4
Department of Computer Science and Engineering, St. Joseph’s Institute of Technology, Chennai, India

E-mail : [email protected], [email protected], [email protected], [email protected]

Abstract- This paper discusses the story of to identify the significant situations where 132 instances offake cash in
type of money that if a sample is given money fraud. 2018, which increments 37% to 181 of every 2019 [3].
Traditional techniques are different as well methods
available for the detection of counterfeit money that on its For this reason, we need to take countermeasures that
color, widths, along with unique currency identification
number on it. Today, the time of present-day of advance
can help prevent this scourge. With the appearance of
computational age-progressed computation techniques, PCs and innovation, it is feasible to make excellent
different AI calculations have been created for picture phony notes that are hard to recognize from genuine
handling that gives close to 100% precision of phony money. notes. Truth be told, many phony fake notes mimic the
Strategies for obtaining and acknowledgment over many security features that are often created with real
calculations incorporate associations, for example, shading, money notes. Therefore, we need to develop new ways
shape, paper width, picture sifting note. This paper proposes to help people identify counterfeit notes in a simple
a fake money acknowledgment technique utilizing K-Nearest way.
Neighbors were trailed by picture handling and further
refinement of boundaries. KNN has a high precision of little
informational collections making it alluring utilized for PC
AI methods that can be utilized to fabricate devices can
discovery work. The banknote picture properties dataset has be useful in this undertaking. We can cause our PCs to
been made with the development of computational and learn examples or elements to help recognize fake and
numerical methodologies , which results in the right genuine cash. In this, the models will utilize ML and
information and data in regards to the substances and picture handling calculations which gauges and learns
elements identified with the money. Information handling their elements . By utilizing these examples and the
and information Extraction is finished by utilizing AI and data that these models can learn, they will actually want
picture calculations handling to get the end-product and to separate new monetary forms as fake or genuine.
exactness. Given the undeniable degree of fake illegal tax
Keywords— Fake Currency, KNN, Image Processing,
avoidance, which is being fostered these days
Currency image detection. determined to bamboozle individuals and adversely
affecting the nation's economy, it is prudent to utilize
I. INTRODUCTION PCs to assist with separating them.

Each significant economy in present-day times needs Hence, primary point is to foster such a model that can
to manage fake cash. A fake money is a kind of money be effectively and viably prepared with the assistance of
that is produced with next to no legitimate important information, and afterward, used to
remuneration by the state or government. Some of its distinguish deceitful obligations with minor order
negative social effects include depreciation of real blunders. Thus, this will be an arrangement model that
money, as well as inflation due to the spread of large will be prepared with the assistance of named
sums of money in the economy. This is the reason why information. This information will contain
governments spend so much money to fight economic characteristics and names for both genuine and phony
warfare. cash bills. Our model will concentrate on these trait
numbers and, with the assistance of related names,
Models incorporate the high-profile outrage that ejected attempt to recognize choice restricts that partition the
in Hungary in 1926, when a couple of individuals were two classes of tests.
captured in the Netherlands while attempting to recover
from the fake French 1000-franc loans produced in
Hungary [1]. According to the U.S. Treasury
Department, an estimated R70 million counterfeit debt
is being distributed in the United States [2]. One of the

Authorized licensed use limited to: Marathwada Mitra Mandal's College of Engg. Downloaded on August 22,2024 at 08:50:11 UTC from IEEE Xplore. Restrictions apply.
 falls into 90%. These recognitions were done before
just with open cv and python anyway over and over
with current profound getting to know techniques
insights had been amassed with the count of hundred
photos in keeping with section and afterward
estimated [6]. Exactness of preparing and testing sets
have been estimated. This brings the chain kind
productivity that lengthens to a bigger expense rather
than various strategies. So, the idea of transferring the
Fig.1 Dataset Description study done. The clamor become additionally caught,
and this was another difficulty because of which a great
K-closest neighbors (KNN) is a calculation where we deal extra development became required. From that
can store every one of the information, and when point onward, a CNN came into the estimation for the
another information comes, it will order it as per which blunder expulsion. Misfortune propensities had been
class it has a place. This implies on the off chance that normally breaking down in regards to preparing
new information shows up then it tends to be simpler training loss (TL) and validation loss (VL).Accuracy
partitioned into a class through KNN calculation [4]. tendencies were generally analyzed via training
Normally, Euclidean distance (ED) is utilized as length accuracy (TA). In 2021, Deep CNN, and followed by
measurements. Then, at that point, it gives a highlight a way of picture processing are the ways to detect a fake
class among the possible neighbor which is near (when note easily. It is now one of most coherent in modern
k is a whole number). days.

Since KNN necessarily not need a offline training class, III.METHODOLOGY

the main calculation is the online 'search' nearby
neighbors k for a given sample test. When we use A database of this problem was obtained from the UCI
different values of k, it may be produced different Machine Learning Repository [7] shown in figure 1. Data
dividing effects, 1-NN is often used as a benchmark is extracted from images taken from real images and
of other dividers as they provide a good separation to constructed as monetary value. To make digital, a modern
many similar problems of patterns. camera regularly used to test print was utilized. The last
pictures are 400x 400 pixels. Because of the focal point
Here, our dataset is not much large in size, so the KNN and item scope of the dark scale pictures with a goal of
algorithm is better suited to the system and we can around 660 dpi were recognized. The Wavelet Transform
easily expect some great scores. The fake disclosure device was utilized to separate elements from pictures.
framework decreases the utilization of fake cash as well
as urges retailers to acknowledge/make cash After wavelet transformed image, the features of the
installments, keeps individuals from spreading fake dataset were observed and it has the given attributes :
notes and guarantees appropriate income in the  Variance
economy.  Skewness
 Kurtosis
II.LITERATURE SURVEY  Image Entropy
 Currency class which is an integer
With the passage of time and increase in the power of
computation, various research and studies related tothis The top four elements continue and state the features of a
field has been done along with various improvement currency note, and the fifth element is a real note label, 1
and techniques [4]. In earlier times, research the facts of the original, and 0 of the dictator. This database
accrued to the detection of counterfeit currency was contains 1372 samples, of which 762 samples are false
with expert pictographic techniques but in those currencies, and 610 samples are real currencies.
information, accuracy visible was good and true due to
easy machine learning (ML) algorithms. KNN In this paper, the goal is to tune a best model so fine
algorithms were popularly used to detect the false such that model that can distinguish new input images
money. Computation power tends to decrease when based on features released after wavelet conversion.The
data length became massive. Later that gadget came all point of focus here is that which techniques we Use to
through to classify the accuracy and acknowledgment further refine the modal if possible. Therefore, we are
rate with a couple of upgrade in AI calculations and working on converted images. We can also measure the
profound dominating ideas [5]. Since the datasets are performance of our model by passing a set of labelled
getting quite enormous, it become easily get confused data that is unknown to our model. We expect to get the
or changed until they are hard to recognize and the rate highest accuracy of our model because the issue is

Authorized licensed use limited to: Marathwada Mitra Mandal's College of Engg. Downloaded on August 22,2024 at 08:50:11 UTC from IEEE Xplore. Restrictions apply.
 related to finance and will cause a huge loss to the SVC :
economy if not properly analysed. SVM is a regulated learning method that forms a
hyperplane or set of hyperplanes in a high-or interminable
We note the following points from this distribution: dimensional space, which can be used for plan, backslide,
 There are no features in this database where the or various endeavors. Regularly, a pleasant fragment is
samplehas a value greater or less than the value in all refined by the hyperplane that has the best distance to the
samples. closest arranging illuminating thing of any class (inferred
 As outliers are not present in our data, we don't need logical edge), since overall the more noteworthy the edge
any logarithmic correction. the lower the theory mess up of the classifier.
 Variance‟ and „skewness‟ are well evenly distributed
inthe dataset. Gradient Boosting Classifier :
 „Kurtosis‟ is a well-distributed distribution, and most Gradient Boosting model is an outfit of frail students to
records have a value of kurtosis less than 3. The such an extent that this total in the end gives an
spread of kurtosis below 3 is said to be platykurtic, exceptionally hearty model. A powerless student is
implying that the spread produces less outside and fundamentally a basic expectation model which is
less continuous irregularities. round dissemination of characterized as one whose presentation is to some extent
kurtosis .It implies that viable pictures ought to have somewhat better than arbitrary possibility. Along these
less commotion. lines, it creates an extremely precise forecast rule by
 Entropy has a distorted distribution, which implies consolidating harsh and decently erroneous standards.
that most samples have high entropy, which also This powerless student or expectation rule is as choice
means that most images have high brightness. trees, which are basically a flowchart of Yes/No inquiries.
 Scaling is necessary to make the features in the same The feeble understudies are added dynamically, one in
range. each resulting advance, and by weighting the observations,
to put more weight on hard to bundle occasions and less
As we find in the dispersed plots, there is a decent on those generally speakingoversaw well.
distinguishableness between the two classes for some
sets of elements. In this way, we can say that the elements Three of the supervised type learning methodologies are
are marvelous in unmistakable in banknote‟ s two classes, chosen for this issue. These calculations are picked to
furthermore we ought to get raised standard accuracy such an extent that their methodologies are on a very basic
with,a suitable assessment shown in figure 2. level, not the same as one another, so we can cover a wide
range of potential methodologies. We think about after
three calculations for investigation, and would contrast
their presentation and our benchmark model.
 Support Vector Classifier (SVC)
 Gradient Boosting Classifier
 K-Nearest Neighbors (KNN) Classifier

Gradient Boosting Classifier will work for the current

issue, as the information is organized such that it is in
suitable structure to pose layered inquiries like in decision
trees. Likewise, with a modest number of highlights, a
helped troupe model ought not take a lot of time during
preparing or testing stage. Along these lines, GBM will
give us a powerful model utilizing a troupe of feeble
students.

KNN Classifier
KNN is a class of sluggish sets which orders a given
informative item by seeing its neighbors and relegating
loads to them so that the nearest neighbors have a more
prominent say in deciding the class. In this, the differences
between points can be Euclidean, Minkowski, and so on
KNN is said to a sluggish student as it doesn't fabricate a
model on preparing information and arranges a point in
test set simply by seeing its closest neighbor.
Fig.2 Scatter plots of continuous features

Authorized licensed use limited to: Marathwada Mitra Mandal's College of Engg. Downloaded on August 22,2024 at 08:50:11 UTC from IEEE Xplore. Restrictions apply.
 A portion of the hyperparameters that KNN controls Standardization
incorporates: From investigating the information, we recognize that we
 n_neighbors: It is number of neighbors to be utilized really want to perform scaling on the four constant
for order highlights to bring them down in scope of 0 to 1. Thusly,
 loads: level of slam every knowledge focuses has. the classifier will treat all highlights similarly. We utilize
 calculation: Algorithm used to register the closest the accompanying recipe to standardize our elements:
neighbors (ball_tree, kd_tree)

For this the steps we will follow areasfollows-

 Preprocess and standardize the informational
collection.
 Divide the information utilizing K-fold cross We perform scaling with the goal that the directed
approval procedure. learning calculation isn't one-sided towards any
component. The initial five examples after standardization
 Preparing the model utilizing the calculations
are repeated here.
examined previously.
 Ascertain the presentation measures.
Implementation
Since our information is looking great after
At first, the informational index was stacked and
standardization, we will now part it into preparing and test
preprocessed. As mentioned within the higher than parts
sets. After, at that point, we will train our proposed model
knowledge standardization was needed that was done
by the preparation set, and afterward assess its exhibition
victimisation the MinMaxScaler within the sklearn
on the test set. Further, at that point, contrast the outcomes
preprocessing library.
and the criterion model.
Benchmark
Splitting data
A benchmark model is depicted to give an edge and
research execution of the strategy acquired by the set up
model. For the Banknote Identification issue, we will
characterize our benchmark to be a guileless classifier,
which groups all cash notes as phony. The exactness of
this gullible classifier will be the quantity of tests of phony
banknotes to the all outnumber of tests in the information.
Our model ought to perform obviously better than the
gullible way to deal with be any commendable for
expected use. Since this gullible forecast model doesn't
consider any data to prove its case, it sets up a benchmark
which let us know that the proposed system is performing
respectably. That being said, utilizing a credulous
expectation would be inconsequential as it will anticipate
all notes as fake and would not distinguish any example as
certified.

We ascertain the precision of the gullible indicator model

as follows:

Procedure
Information Preprocessing
Information Preprocessing is important to bring
information into great shape, eliminate any anomalies, or
Fig.3 Performance metrics for three models
alter any qualities, prior to taking care of it into a
calculation. From the information representation area, we
The explanation we do by divising is to test the prepared
noted after focuses in regards to our information.
model over examples which cannot be visible by it.
 The dataset doesn't have any missing worth. Thusly, we ensure that our proposed algorithm has
 All of the information features are reliable and have removed grouping designs from the preparation tests and
assorted reach, so we need to perform normalization does not remember them. For the given issue, to divide our
to scale all components in the extent of 0-1. information, we utilize a capacity train_test_split
accessible in cross_validation module of sklearn library.
This capacity completes two errands which are sufficient

Authorized licensed use limited to: Marathwada Mitra Mandal's College of Engg. Downloaded on August 22,2024 at 08:50:11 UTC from IEEE Xplore. Restrictions apply.
 to present scenario as given in figure 3. money notes , i.e., both got a review of 1.0.
 It rearranges the dataset with the goal that both  GBM and KNN have the most minimal bogus up-
preparing and sets of test have almost equivalent sides, and in this manner have the most noteworthy
number of tests from the two classes. accuracy.
 Subsequent to rearranging, it plays out the split. We  KNN registers a precision of 0.9982, and F-score of
also indicate what part of absolute information to be 0.9992 in the set of tests. It orders all tests effectively
remembered for the preparation or test set. aside from only one bogus positive. It is particularly
great.
Subsequent to parting, the capacity gives four records,  KNN Classifier outflanks above models calculations
including input highlights of the preparation set, input on both preparing and set of tests, of every sizes of
elements of the test set, target names for the preparation preparing tests.
set, and target marks for the test set. We will divide to  From the representations and the assessment
such an extent that the preparation set has 60% of tests, measurements, it is noticeable that KNN is a
while the testing set should get 40% of tests. reasonable champ, however different calculations
need improvementsto perform similarly well.
Making Training and Prediction of the Pipeline  However slope supporting sets aside some effort for
We characterize a strategy called 'train_split' in which preparing, and KNN sets aside duration of testing,
given input as info the accompanying boundaries: student, the general time taken is tiny, of the request for part
sample_size, X_train, y_train, X_test, y_test. It returns the of a second.
exactness and F-beta score on preparing and test set. The  Such great outcomes were expected from the
capacity suits the 'student' on the preparation information dispersed plots that has created earlier.
of size characterized by 'sample_size', and processes the
time it taken on preparing. And after that it forecasts on Since K-Nearest Neighbor Classifier is better than the
the test set, and 300 examples of preparing set, and other listed model here, and ordered all phony cash notes
register the time spent on expectation. At last, we register effectively, it is the calculation of our decision for tackling
the exactness and f-beta scores on training as (300 this issue.
examples) and test set. We likewise process disarray lattice
to improve the representation of the forecasts. Refinement
However here we get some great outcomes with K-NN
The training will be performed on the above models, Classifier , there is no such additional refinement required,
each with 0.05x, 0.2x, and 1x of the preparation yet we will in any case play out a Grid Search. The
information as test size, so it tends to be perceived how boundaries that we might want to advance for KNN
execution changes with differing numbers of train set. The calculation are:
structured presentations portraying the exactness and f-  'n_neighbors' : We will take six qualities [1, 2, 3, 4,
scores of the above models on preparing and test sets are 5, 6]
portrayed in the result analysis section.  'loads': weight work utilized in forecast. Two of the
qualities ['uniform', 'distance']. 'uniform' esteem
IV.RESULT ANALYSIS gauges many points in the local similarly, while
'distance' esteem puts more impact on nearer
Subsequent to preparing the model utilizing the neighbors.
informational index and the calculations depicted above,  'calculation': calculation used to process the closest
we tried them to perceive how they are efficient. neighbors. We accept potential qualities as
Ascertaining the presentation measures depicted above, we ['ball_tree','kd_tree']
had the option to acquire an agreement of how each of
these calculation performed. The following are the With these boundaries to improve for, we play out a
exactness, accuracy, and f-scores for every calculation. It GridSearchCV to track down the worth mix of boundaries.
results in the accompanying scores when the whole train
set is utilized for preparing. Confusion matrix of Feature Importance
expectation of test information is additionally given At the end we‟ll try out figuring which highlights give the
underneath to each of the three classifiers:Following most prescient power. By zeroing in on the connection
focuses are noted from the above perception and between a couple of essential elements and the objective
measurements: name we work on our comprehension of the peculiarity.
 Every one of the three calculations perform However, we just have a little arrangement of highlights,
extraordinarily well on the set of tests, that is needed and we have some unpleasant thought from the dispersed
from the system that distinguishes counterfeit cash plots too, we will presently measure this thought of
notes. deciding element significance. For that purpose , we pick a
 SVC and KNN calculations recognize all phony scikit-learn classifier (e.g., AdaBoost, arbitrary

Authorized licensed use limited to: Marathwada Mitra Mandal's College of Engg. Downloaded on August 22,2024 at 08:50:11 UTC from IEEE Xplore. Restrictions apply.
 timberlands) that has a 'feature_importances' The after effects of the model is reliable on many
characteristic, that‟s a capacity that positions the records. Initially, the test set was unique in relation to
significance of elements as per the picked classifier. We preparing set. Along these lines, the model was assessed
pick two calculations, viz., Random Forests and Gradient on inconspicuous examples in the stage of testing.
Boosting to rank the elements and both give comparable Besides, we likewise did network search with cross-
outcomes. approved randomized folds to discredit any odds of
overfitting, which yields comparable outcomes. Then, we
Every one of the three classifiers that we talked about have comparative outcomes with different models too. In
created great precision and f-beta, scores, yet KNN this way, indeed, the outcomes can be trusted.
classifier stood apart from these other techniques. Thus,
this is our last decision of model as an answer for the At last, we accept that the model is sensible and lines up
stated issue. We utilize the following measures to choose with our assumptions for a decent arrangement as it
KNN as the last model. performs extraordinarily well and is very powerful. The
model is likewise performing extremely quick and
Initially, the KNN is better in terms of precision and f-beta delivering great arrangement brings about part of a
scores than SVC and GBM, in both preparing and test sets. second.
The exactness and f-beta score on preparing set is 1.0, and
onset of tests 0.9982 and 0.9992, individually. In our test By the way, we additionally register the prescient force of
set, it grouped all phony notes accurately and falsely find each component, utilizing two calculations, one of them is
only one real note. Random Forests and other one is GBC . These calculations
have 'feature_importances_' trait that allows a number to
Besides, the model beats SVC and GBM reliably on each component in the scope of 0-1, with the end goal
different preparing set sizes. In any event, when preparing that their aggregate is 1. The results are displayed in Fig 9.
set was 5% of its genuine size, expectation precision was We can see that initial three highlights, viz., fluctuation,
over 90% on both preparing and test information. This skewness and kurtosis have the highest prescient power,
connotes the heartiness of the calculation. but the impact of entropy is least.

Thirdly, the outcomes have been so surprising with KNN, V.CONCLUSION

that regardless, adjusting the restrictions of model to
organize search made same outcomes. A portrayal of However we have outstandingly great outcomes, but there
results is shown. On review the result of the matrix search, is a room that indicates possibilities of progress, as in a
we observe that the enhanced model has boundary monetary application like this, we generally intend to be
'n_neighbors' = 1. 'n_neighbors' characterizes the quantity 100% exact. Perhaps, It is further increased by considering
of nearest neighbors that cast a ballot in the expectation a bigger dataset which catches additional distinctive
result. However, at first, we choosen the worth of examples, and investigating different calculations like
'n_neighbors' to be 5. This roused us to see what number Neural Networks, which we didn't contact in this
of neighbors means for the presentation of KNN with undertaking. One more method for taking care of this issue
regards to this issue. The outcomes are portrayed. We can be using profound learning. We can prepare a
see that the exactness and F-beta score are the same for up convolutional neural organization, straight over the caught
to ten nearest neighbors. This again features the vigor of pictures, to construct an older model. Another chance
KNN Classifier for the same issue. which can be investigated is to utilize pre-prepared
profound neural organizations like AlexNet or VGGNet. I
All in all, what worth of 'n_neighbors' to be considered firmly feel that these techniques ought to be investigated as
in the last arrangement model? Worth of 1 appears to be potential arrangement applicant.
extremely basic, and makes our model exceptionally
touchy to the info information. A solitary loud point or REFERENCES
exception might change the genuine mark of certain
examples. Worth of at least 5 makes the model minimal [1] https://en.wikipedia.org/wiki/Counterfeit_money
complex. In this way, we ought to go with the worth of https://en.wikipedia.org/wiki/Counterfeit_United_State_
Currency
one or the other 3 or 4. In any case, 4 is a considerably [2] S. Arya and M. Sasikumar, “Fake Currency Detection,” 2019
input, so we can continue with 3 as the worth of Int. Conf. Recent Adv. Energy- Efficient Comput. Commun.
'n_neighbors' in our last model of KNN classifier. ICRAECC 2019, pp. 2019–2022, 2019,
doi:10.1109/ICRAECC43874.2019.8994968.
[3] D. V. Kapare, S. Lokhande, and S. Kale, “Automatic Cash
Thus, it is observed that the model is very strong as it Deposite Machine with Currency Detection Using Fluorescent
gives shockingly comparative outcomes on different And UV Light,” vol. 3, pp. 309–311, 2013.
preparing set sizes just as on various upsides of its [4] T. Agasti, G. Burand, P. Wade, and P. Chitra, “Fake currency
boundary 'n_neighbors'. detection using image processing,” IOP Conf. Ser. Mater. Sci.
Eng., vol. 263, no. 5, pp. 88– 93, 2017, doi:10.1088/1757-

Authorized licensed use limited to: Marathwada Mitra Mandal's College of Engg. Downloaded on August 22,2024 at 08:50:11 UTC from IEEE Xplore. Restrictions apply.
 899X/263/5/052047. [11] Selvakumar, K., & Sairamesh, L. (2021). User query-based
[5] P. P. Binod Prasad Yadav, C. S. Pat il, R. R. Karhe,"An automatic text summarization of web documents using
automatic recognition of fake Indian paper currency note using ontology. In International Conference on Communication,
MATLAB," Certif. Int. J. Eng. Sci. Innov. Technol., vol. 9001, Computing and Electronics Systems: Proceedings of ICCCES
no.4, pp. 2319– 5967, 2008. 2020 (pp. 593-599). Springer Singapore.M. A. Gaikwad, V. V
[6] S. Adhikari, S. Thapa, and B. K. Shah,“Oversampling based Bhosle, and V. D. Pat il, “Automatic Indian New Fake Currency
Classifiers for Categorization of Radar Returns from the Detection Technique,” Int. J. Eng. Res. Technol., vol. 6, no. 11,
Ionosphere," Proc. Int. Conf. Electron. Sustain. Commun. Syst. pp. 84–87, 2017.
ICESC 2020, no. Icesc, pp. 975–978, 2020, doi: [12] SaiRamesh, L., Ashok, E., Sabena, S., & Ayyasamy, A. (2020).
10.1109/ICESC48915.2020.9155833. Credit Card Fraud Detectıon in Retaıl Shopping Using
[7] A. Kumar and A. Kumar, “Dog Breed Classifier for Facial Reinforcement Learning. New Trends in Computational Vision
Recognition using Convolutional Neural Networks,” pp. 508– and Bio-inspired Computing: Selected works presented at the
513, 2020. ICCVBIC 2018, Coimbatore, India, 1541-1549.A. Ghimire, S.
[8] M. Haider Ali, “Thesis Report on Fake Currency Detection Thapa, A. K. Jha, S. Adhikari, and A. Kumar, “Accelerating
using Image Processing Method,” Akiful Mohaimin Rifat Islam business growth with big data and artificial intelligence,” Proc.
Shahriar Chowdhury, no.13301148, pp. 1–38, 1330. 4th Int. Conf. IoT Soc. Mobile, Anal. Cloud, ISMAC 2020, pp.
[9] M. N. Shende and P. P. Patil, “A Review on Fake Currency 441–448, 2020, doi: 10.1109/ISMAC49090.2020.9243318
Detection using Image Processing,” Int. J. Futur. Revolut.
Comput. Sci. Commun. Eng., vol. 4, no. 1, pp. 391–393, 2018.
[10] Sulthana, R., Jaiswal, A., Supraja, P., & Sairamesh, L. (2023,
January). Customer Segmentation using Machine Learning.
In 2023 Third International Conference on Advances in
Electrical, Computing, Communication and Sustainable
Technologies (ICAECT) (pp. 1-5). IEEE.

Authorized licensed use limited to: Marathwada Mitra Mandal's College of Engg. Downloaded on August 22,2024 at 08:50:11 UTC from IEEE Xplore. Restrictions apply.