ISSN 2347 - 3983

Sumit Kumar et al., International Journal of Emerging

VolumeTrends
9. No.in8,Engineering
August 2021Research, 9(8), August 2021, 1046 – 1050
International Journal of Emerging Trends in Engineering Research
Available Online at http://www.warse.org/IJETER/static/pdf/file/ijeter01982021.pdf
https://doi.org/10.30534/ijeter/2021/01982021

A Review: Machine Learning Approach and Deep Learning

Approach for Fake News Detection
Sumit Kumar1, Jyoti Tiwari 2
1
M.E. Computer Engineering Sgsits Indore, India, bhadouriasumit27@gmail.com
2
Assistant Professor Comp. Tech & Appl. Dept. Sgsits Indore, India, jyotimona23@gmail.com


ABSTRACT
finding the best feature set for a particular problem on its own.
[1].
Increasing number of social media platforms, emerging
new technologies, and population growth which results in the
2. MACHINE LEARNING ALGORITHM APPROACH
rate of using social media has increased rapidly. With an
increasing number of users on online platforms comes to a
variety of problems like fake news. The extensive growth of The aim of machine learning methods is to automate the
fake news on social media can have a serious impact on the learning process of computers based on evidence. It splits the
real world and became a cause of concern for net users and data into training and research, with the training data being
governments all over the world. Distinguishing between real used to learn the model and the testing data is used to validate
news and fake news becoming more challenging. The amount the data. It is divided into three categories: supervised,
of fake news has become a disguise. In this paper, we have semi-supervised, and unsupervised learning. For learning the
done a survey on detection techniques for fake news using model, supervised machine learning methods use labeled
Algorithms and Deep learning techniques. We have results. Unlabeled data, on the other hand, is analyzed using
compared machine learning algorithms like Naïve-Bayes, unsupervised techniques.
Decision tree, SVM, Adaboost, etc. Comparing the accuracy
2.1 Naïve bayes classifier
Key words: Convolutional neural network, support vector Naive Bayes is a probabilistic machine learning algorithm,
machine, LSTM. and this is a probabilistic classifier that uses the max
posteriori decision rule to render classifications. The Bayes
1. INTRODUCTION
theorem is used in classification activities, and the Naive
Bayes classifier is derived from it [15]. This classifier is a
Fake news has become one of the biggest threats to versatile and common conventional algorithm for practical
journalism. It has weekend people trust on News agencies and uses. Users get a short response from these real-world apps.
government. The method of classifying news stories as false This algorithm is used in email spam filtering and sentiment
or real is known as fake news identification. To accomplish analysis, among other things. This is one of the most basic
this, it must first extract specific and important features from classification problems, and it requires large datasets.
the article's text and then use a classifier to separate fake news ble, approximately 8 to 12 point type.
from genuine news... To overcome this by using some
websites, to overcome this by using some websites, tools, and P(I/J)=(P(J/I)*P(I))/P(J)
platforms were introduced to detect the differentiation of real Posterier=(Liklihood*prior)/Evidence
news and fake news. The fakesters create a piece of news that
makes the readers' minds desperate to learn what is there in P(I): Prior Probability of class
the news by seeing the title of particular news. That desperate P(J|I): Likelihood, the probability of predictor given class
minds are giving a chance to the fakesters to create such P(J): Prior probability of predictor
unhealthy news. Even though fake news detection is not P(I|J): Posterior probability off class I was given the predictor
completely ruling the world but a lot of interest is being J
generated differentiation of real news and fake news. Many
researchers have worked on this topic. Some of them used 2.2 K-Nearest Neighbor
Traditional classification algorithms and others used Deep The supervised machine learning algorithm known as the
learning approaches. Deep learning has the advantage of k-nearest neighbor algorithm solves classification and
having no handcrafting of rules and/or functions, instead of regression problems. Let's take a look at some labeled points,
measure the distance between them, arrange the distances in
decreasing order, and mark the first point as "k." The

1046
Sumit Kumar et al., International Journal of Emerging Trends in Engineering Research, 9(8), August 2021, 1046 – 1050

regression and classification models' mean and mode are information is reduced and the output is provided by the
returned by the K marks. This is a delicate model to deal with pooling layer. CNN looks for dormant personalities in news
and the outputs of classification problems are discrete values stories. The data size and qualified data are the most common
[14]. The K nearest model identifies patterns and senses uses of this model. Based on efficiency and speed, the CNN
intrusions. model is considered.

2.3 Support Vector Machine 3.2 Long-Short Term Memory

A supervised learning algorithm is SVM. This is one of the A long short-term memory is a persistent recurrent neural
best-fit algorithms as compared to other algorithms. To find network (LSTM). In LSTMs, individual hidden drives are
false news, SVM employs computational linguistics. This used, and their typical action is to remember inputs for a long
algorithm is used to transform learning models that are only time. A memory cell, also known as an accumulator or a gated
intended for single use. The training sets and data sets are also leaky neuron, is a system that works as an accumulator or a
normalized. It has a wide variety of implementations in the gated leaky neuron and has a weight of 1. In other words, it
qualified set. A commonly used classification algorithm is the repeats itself.
help vector machine. Often used for collecting vast amounts
of data from images with a small number of dimensions.
3.3 Feedforward Neural Network
2.4 Decision Tree Feedforward NNs are the first and, in many ways, the most
A decision tree is a flowchart representation of a supervised basic kind of artificial neural network. In this network, data
learning model. It's a non-parametric approach for only flows in one direction: forward from the input nodes to
classification and regression that's simple and easy to use. We the output nodes, passing through any hidden nodes (if any).
partition the root set into subsets using a decision tree In the network, there are no cycles or loops. Feedforward
dependent on attribute value tests. Recursively, the division networks (also known as associative networks) are made up of
process is repeated on each subset. When all of the subsets at various categories of components, such as binary
a node have the same variable, the recursion is complete. The McCulloch–Pitts neurons (Anderson, 2006), with the
decision tree with decision nodes and leaf nodes is the perceptron becoming the most basic. In the case of
end product. A decision node has two or three nodes, with the backpropagation of errors, continuous neurons with sigmoid
leaf node representing either grouping or decision. It is activation functions are commonly used.
capable of handling both categorical and numerical results.
3.4 Bidirectional LSTM concatenated model
2.5 Random Forest Using Google's pre-trained model, the headline and article
Many random decision trees make up a random forest. The texts are translated into two different embedding layers. The
trees have two forms of randomness built-in. To begin, each two embedding vectors' results are concatenated to feed the
tree is constructed from a random sample of the original data. construct. The model is made up of two CNN layers, each
Second, a subset of features is randomly chosen at each tree with 32 and 64 filters. Until moving it via a Bidirectional
node to produce the best break. LSTM layer with 100 memory units, the CNN's are
accompanied by max-pooling to prevent over-fitting5. The
bidirectional LSTM's output was routed through three dense
3. DEEP LEARNING ALGORITHM APPROACH layers (512, 128, 4 units) separated by a single dropout.

Deep Neural Networks (DNNs) were introduced as a way to

replicate the human brain's pattern recognition abilities. A 4. LITERATURE REVIEW
DNN is a neural network with input, output, and a single
hidden layer that is made up of networks. Finds mathematical In the field of machine learning, a large number of
manipulations to turn on non-linear or linear interactions in experiments have been performed for the identification of
the input and output. false news. B. M. Amine, A. Drif and S. Giordano [2] the
approach that is taken to solve the problem is through a
3.1 Convolutional Neural Network convolutional neural network. Compared to other methods
Convolutional neural networks are a network of neurons that based on CNN where a single CNN-based model is
are linked in layers that are used to take inputs and outputs. constructed using a text-only approach means only text
CNN is a feed-forward network model that can be used to metadata is used. But using more than text metadata by using
recognize objects and analyze images. The convolution layer, another CNN show significant improvement for fake news
detector layer, and pooling layer are the three levels that make detection.
up the neural network that performs the functions. The
convolutional layer's task is to build a complicated function R. K. Kaliyar et al. [3] the dataset they have used is a
chart. The nonlinear components of function maps are multiclass fake news dataset which is taken from the Fake
highlighted by the detector layer. The predecessor
1047
Sumit Kumar et al., International Journal of Emerging Trends in Engineering Research, 9(8), August 2021, 1046 – 1050

news challenge-based fake news dataset. Benchmark findings Authors B. M. J. Lin, V. M. E.

are available in the latest analysis focused on two groups' Amine, G. Krešňák Qawasm
datasets, classifying news as false or real? There is minimal A. Drif Trembla ová, M. eh, M.
literature on multiclass estimation. There were many contents and S. y-Taylor Sarnovs Tawalbe
and context-related issues with the binary class dataset. So, Giordan , G. ký and h and M.
they have used a multiclass dataset. They utilized gradient o [2] Mou, D. P. Butka Abdulla
boosting with an advanced parameter that fits best for You and [9] h [10]
multi-class fake news datasets. K. Lee
[4]
J. Lin et al. [4] have built a Long-term short-term memory
(LSTM) based machine learning model with the Paper Merging Detectin Deep Automat
attention-based model as LSTM suffers from the issue of deep g Fake learning ic
long-range dependence and the encoder-decoder network learning News methods Identific
output degrades rapidly as the sentence length increases. To model Articles for Fake ation of
counter the issue attention-based LSTM is used. Other for fake News Fake
machine learning models such as Random Forest, XGBoost, news detectio News
etc. are used as the performance measure. detectio n Using
n Deep
Ye-Chan-Ahn, Chang-Sung Jeong [6] a method is designed Learnin
to counter fake news using advantage and disadvantage of g
LSTM and CNN as LSTM takes a longer time to learn and is
vulnerable to parallelism and CNN which require a lot of Metrics ACC:96 ACC:91 ACC:91 ACC:85
computation to connect two input and output that are located % % % %
far away from position. So, they have a transformer network PREC:9 PREC:9 PREC:8 PREC:7
that uses both self-attention and feed-forward mechanisms. 8% 0% 5% 7%
REC:94 REC:93 REC:99 REC:75
Liang Wu and Huan Liu [7] States that the classification of % % % %
messages was wont to find the messages spreading in social Paramet Accurac Accurac Accurac Accurac
media which are not appropriate, trustworthy, etc. And er y, y, y, y,
proposed Trace Miner to classify messages in social media, Precisio Precisio Precisio Precisio
with this an end-to-end LSTM-RNN’s Classification model n, n, n, n,
was also included. Trace Miner has given good results on the Recall Recall Recall Recall
insufficient datasets. The longer-term work goes to be done on
network mining tasks using Trace Miner. Advanta At once They Can Solves
ges multiple could represen the
Ahmed, Hadeer, Issa Traore, and Sherif Saad [8] two classific solve the t more problem
different attribute extraction techniques and six different ations problem complex of fixed
classification techniques for machine learning have been are of function sequenc
investigated and researched. Compared to state-of-the-art done, vanishin s very e to
techniques, they need to illustrate the efficiency of certain locates g easily, It sequenc
models and demonstrate better performance. The most objects gradient will e
significant aim of this research article was to automatically in an s. learn predictio
identify fake news material. For their study, they need to image more n.
recognize false feedback and fake news results. complex
function
Table 1 :Work Done In Deep Learning Classification models. s.
Disadva Comput require a Prone to Since
Algori CNN LSTM Feedfor Bidirect ntages ational lot of overfitti BiLSTM
thms ward ional cost is resource ng, has
neural LSTM expensiv s and Vanishi double
network Concate e, time to ng and LSTM
nated require get Explodi cells so
model lot of trained ng it is
training gradient costly
data problem.

1048
Sumit Kumar et al., International Journal of Emerging Trends in Engineering Research, 9(8), August 2021, 1046 – 1050

Table 2 Work Done in Machine Learning Classification Models.

Advan Easily Perfor Simple Interpr Powerf
tages implem mant, Implem etabilit ul and
Algorit Naives SVM KN Decisi Rando ented not entatio y, accurat
hms -Bayes on m , biased n, no need e, good
Tree Forest require by classes for perfor
s outliers don't feature mance
Author M. P. B. P. J. Lin, S. Lyu M. P. less , not have to scaling, on
s Granik Reddy, G. and D. K. trainin sensitiv be works many
and V. M. P. Trembl C. -T. Reddy, g e to Linearl on both proble
Mesyu K. ay-Tay Lo [11] K. M. period over y linear / ms,
ra [12] Reddy, lor, G. Mehat to test fitting. separab non – includi
G. V. Mou, a , P. the le, linear ng non
M. D. You B. P. data, Robust proble –
Reddy and K. Reddy fast with ms. linear.
and K. Lee [4] and G. decisio regard
M. V. M. n to
Mehat Reddy making search
a [13] [13] . space.

Disad Loss of Not Expens Poor No

Paper Fake Fake Detecti Fake Fake vantag accurac appropr ive results interpr
news Data ng News Data es y iate for testing on very etabilit
detecti Analys Fake Detecti Analys non-lin of small y, over
on is and News on by is and ear each dataset fitting
using Detecti Article Decisi Detecti proble Instanc s, can
naive on s on on ms, not e, over easily
Bayes Using Tree Using the best Sensiti fitting occur,
classifi Ensem Ensem choice veness can need to
er bled bled for to easily choose
Hybrid Hybrid large noise, occur. the
Algorit Algorit number Sensiti number
hm hm of veness of trees
feature to very manual
Metric ACC:7 ACC:8 ACC:7 ACC:9 ACC:8
s. Unbala ly.
s 5.40% 9% 9% 5% 6%
nced
PREC: PREC: PREC: PREC: PREC:
data.
71% 85% 75% 92% 83%
REC:7 REC:9 REC:7 REC:9 REC:8
3% 3% 9% 0% 9%

Param Accura Accura Accura Accura Accura

eter cy, cy, cy, cy, cy,
Precisi Precisi Precisi Precisi Precisi 5. ACKNOWLEDGMENT
on, on, on, on, on,
Recall Recall Recall Recall Recall
My thanks to the Jyoti Tiwari mam who have contributed
towards the development of the paper.

6. CONCLUSION

A comparison of different methods identified by various

researchers to deal with the issue of fake news identification
has been done in this article. The challenge of detecting fake
news in this age is causing a lot of problems. There has been a
lot of studies undertaken to find ways to prevent and reduce
the issue of fake news. Various methods have been outlined in
this paper to do.

1049
Sumit Kumar et al., International Journal of Emerging Trends in Engineering Research, 9(8), August 2021, 1046 – 1050

REFERENCES [10] E. Qawasmeh, M. Tawalbeh and M. Abdullah,

"Automatic Identification of Fake News Using
[1] A. Roy, K. Basak, A. Ekbal and P. Bhattacharyya, Deep Learning," 2019 Sixth International
"A Deep Ensemble Framework for Fake News Conference on Social Networks Analysis,
detection and classification," arXiv preprint Management and Security (SNAMS), Granada,
arXiv:1811.04670, 2018. Spain, 2019, pp. 383-388, doi:
10.1109/SNAMS.2019.8931873.
[2] B. M. Amine, A. Drif and S. Giordano, "Merging
deep learning model for fake news detection," [11] S. Lyu and D. C. -T. Lo, "Fake News Detection by
2019 International Conference on Advanced Decision Tree," 2020 SoutheastCon, Raleigh, NC,
Electrical Engineering (ICAEE), Algiers, Algeria, USA, 2020, pp. 1-2, doi:
2019, pp. 1-4, doi: 10.1109/SoutheastCon44009.2020.9249688.
10.1109/ICAEE47123.2019.9015097. [12] M. Granik and V. Mesyura, "Fake news detection
[3] R. K. Kaliyar, A. Goswami and P. Narang, using naive Bayes classifier," 2017 IEEE First
"Multiclass Fake News Detection using Ensemble Ukraine Conference on Electrical and Computer
Machine Learning," 2019 IEEE 9th International Engineering (UKRCON), Kyiv, UKraine, 2017, pp.
Conference on Advanced Computing (IACC), 900-903, doi: 10.1109/UKRCON.2017.8100379.
Tiruchirappalli, India, 2019, pp. 103-107, doi: [13] P. B. P. Reddy, M. P. K. Reddy, G. V. M. Reddy and
10.1109/IACC48062.2019.8971579. K. M. Mehata, "Fake Data Analysis and Detection
[4] J. Lin, G. Tremblay-Taylor, G. Mou, D. You and K. Using Ensembled Hybrid Algorithm," 2019 3rd
Lee, "Detecting Fake News Articles," 2019 IEEE International Conference on Computing
International Conference on Big Data (Big Data), Methodologies and Communication (ICCMC),
Los Angeles, CA, USA, 2019, pp. 3021-3025, doi: Erode, India, 2019, pp. 890-897, doi:
10.1109/BigData47090.2019.9005980. 10.1109/ICCMC.2019.8819741.
[5] P. Shah and Z. Kobti, "Multimodal fake news [14] A. Kesarwani, S. S. Chauhan and A. R. Nair, "Fake
detection using a Cultural Algorithm with News Detection on Social Media using K-Nearest
situational and normative knowledge," 2020 Neighbor Classifier," 2020 International
International Conference on Advanced Computing Conference on Advances in Computing and
(IACC), Glasgow, United Kingdom, 2020, pp. 1-7, Communication Engineering (ICACCE), Las
doi: 10.1109/CEC48606.2020.9185643. Vegas, NV, USA, 2020, pp. 1-4, doi:
10.1109/ICACCE49060.2020.9154997.
[6] Y. Ahn and C. Jeong, "Natural Language Contents
Evaluation System for Detecting Fake News using [15] Anurag Busha 1 , Manoj Prakash.P 2 , Pelash
Deep Learning," 2019 16th International Joint Choudhary 3 , Vakeesh Kanna. T. I 4 “Sentimental
Conference on Computer Science and Software Analysis” International Journal of Emerging
Engineering (JCSSE), Chonburi, Thailand, Technologies in Engineering Research (IJETER)
2019,pp. 289-292, doi: Volume 6, Issue 4, April (2018).
10.1109/JCSSE.2019.8864171.
[7] Wu, L., & Liu, H. (2018). “Tracing Fake-News
Footprints”. Proceedings of the Eleventh ACM
International Conference on Web Search and Data
Mining - WSDM ’18.
doi:10.1145/3159652.3159677.
[8] Ahmed, Hadeer, Issa Traore, and Sherif Saad.
”Detecting opinion spams and fake news using
text classification.” Security and Privacy 1, no. 1
(2018): e9
[9] V. M. Krešňáková, M. Sarnovský and P. Butka,
"Deep learning methods for Fake News
detection," 2019 IEEE 19th International
Symposium on Computational Intelligence and
Informatics and 7th IEEE International Conference
on Recent Achievements in Mechatronics,
Automation, Computer Sciences and Robotics
(CINTI-MACRo), Szeged, Hungary, 2019, pp.
000143-000148, doi:
10.1109/CINTI-MACRo49179.2019.9105317.

1050