Leveraging Machine Learning for Fake News Detection

Elio Masciari, Vincenzo Moscato, Antonio Picariello and Giancarlo Sperl

University Federico II, Naples, Italy

Keywords:

Fake News Detection, Machine Learning.

Abstract:

The uncontrolled growth of fake news creation and dissemination we observed in recent years causes contin-

uous threats to democracy, justice, and public trust. This problem has signiﬁcantly driven the effort of both

academia and industries for developing more accurate fake news detection strategies. Early detection of fake

news is crucial, however the availability of information about news propagation is limited. Moreover, it has

been shown that people tend to believe more fake news due to their features (Vosoughi et al., 2018). In this

paper, we present our complete framework for fake news detection and we discuss in detail a solution based on

machine learning. Our experiments conducted on two well-known and widely used real-world datasets sug-

gest that our settings can outperform the state-of-the-art approaches and allows fake news accurate detection,

even in the case of limited content information.

1 INTRODUCTION

Social media are nowadays the main medium for

large-scale information sharing and communication

and they can be considered the main drivers of the Big

Data revolution we observed in recent years(Agrawal

et al., 2012). Unfortunately, due to malicious user

having fraudulent goals fake news on social media are

growing quickly both in volume and their potential

inﬂuence thus leading to very negative social effects.

In this respect, identifying and moderating fake news

is a quite challenging problem. Indeed, ﬁghting fake

news in order to stem their extremely negative effects

on individuals and society is crucial in many real life

scenarios. Therefore, fake news detection on social

media has recently become an hot research topic both

for academia and industry.

Fake news detection dates back long time

ago(Zhou et al., 2019a) as journalist and scientists

always fought against misinformation during human

history. Unfortunately, the pervasive use of internet

for communication allows for a quicker and wider

spread of false information. Indeed, the term fake

news has grown in popularity in recent years, espe-

cially after the 2016 United States elections but there

is still no standard deﬁnition of fake news (Shu et al.,

2017a).

Aside the deﬁnition that can be found in literature,

one of the most well accepted deﬁnition of fake news

is the following: Fake news is a news article that is

intentionally and veriﬁable false and could mislead

readers (Allcott and Gentzkow, 2017). There are two

key features of this deﬁnition: authenticity and intent.

First, fake news includes false information that can

be veriﬁed as such. Second, fake news is created with

dishonest intention to mislead consumers(Shu et al.,

2017b).

The content of fake news exhibits heterogeneous

topics, styles and media platforms, it aims to mys-

tify truth by diverse linguistic styles while insulting

true news. Fake news are generally related to newly

emerging, time-critical events, which may not have

been properly veriﬁed by existing knowledge bases

due to the lack of conﬁrmed evidence or claims. Thus,

fake news detection on social media poses peculiar

challenges due to the inherent nature of social net-

works that requires both the analysis of their con-

tent (Potthast et al., 2017; Guo et al., 2019; Masood

and Aker, 2018; Masciari, 2012) and their social con-

text(Shu et al., 2019; Cassavia et al., 2017; Masciari,

2012).

Indeed, as mentioned above fake news are written

on purpose to deceive readers to believe false infor-

mation. For this reason, it is quite difﬁcult to detect

a fake news analysing only the news content(Shabani

and Sokhn, 2018). Therefore, we should take into ac-

count auxiliary information, such as user social en-

gagement on social media to improve the detection

accuracy. Unfortunately, the usage of auxiliary in-

formation is a non-trivial task as users social engage-

Masciari, E., Moscato, V., Picariello, A. and Sperlì, G.

Leveraging Machine Learning for Fake News Detection.

DOI: 10.5220/0009767401510157

In Proceedings of the 9th International Conference on Data Science, Technology and Applications (DATA 2020), pages 151-157

ISBN: 978-989-758-440-4

151

ments with fake news produce data that are big, noisy,

unstructured and incomplete.

Moreover, the diffusion models for fake news

changed deeply in recent years. Indeed, as men-

tioned above, some decades ago, the only medium

for information spreading were newspapers and ra-

dio/television but recently, the phenomenon of fakes

news generation and diffusion take advantage of the

internet pervasive diffusion and in particular of social

media quick pick approach to news spreading. More

in detail, user consumption behaviours have been af-

fected by the inherent nature of these social media

platforms: 1) They are more pervasive and less expen-

sive when compared to traditional news media, such

as television or newspapers; 2) It is easier to share,

comment on, and discuss the news with friends and

followers on social media by overcoming geographi-

cal and social barrier.

Despite the above mentioned advantages of so-

cial media news sharing, there are many draw-

backs, requiring also an suitable pre-processing

phase(Mezzanzanica et al., 2015; Boselli et al., 2018).

First of all, the quality of news on social media is

lower than traditional news organizations due to the

lower control of information sources. Moreover, since

it is cheaper to provide news online and much faster

and easier to spread through social media, larger vol-

umes of fake news are produced online for a variety

of purposes, such as political gain and unfair compe-

tition to cite a few.

Fake News Examples. Some well known exam-

ples of fake news across history are mentioned be-

low: a) During the second and third centuries AD,

false rumours were spread about Christians claiming

that they engaged in ritual cannibalism and incest

; b)

In 1835 The New York Sun published articles about

a real-life astronomer and a made-up colleague who,

according to the hoax, had observed bizarre life on

the moon

; c) More recently we can cite some news

like, Paul Horner, was behind the widespread hoax

that he was the grafﬁti artist Banksy and had been ar-

rested; a man has been honored for stopping a rob-

bery in a diner by quoting Pulp Fiction; and ﬁnally

the great impact of fake news on the 2016 U.S. presi-

dential election, according to CBS News

Furthermore, in 2018 BuzzFeed News compiled

a list of 50 most viral false stories on Facebook and

measured their total engagement on the platform. And

https://en.wikipedia.org/wiki/Fake news

http://www.snelgraphix.net/the-

snelgraphix-designing-minds-

blog/tag/google+I%E2%80%99m+feeling+stellar

https://www.businessinsider.com/banksy-arrest-hoax-

2013-2

in spite of a prediction from Facebook’s top anti-

misinformation product manager that these articles

would see a decline in engagement in 2018, the top-

performing hoaxes generated roughly 22 million to-

tal shares, reactions, and comments on Facebook be-

tween Jan. 1 and Dec. 9, 2018.

Psychological Aspects behind Fake News. The in-

ﬂuential power of fake news has been explained by

several psychological theories. Fake news mainly tar-

gets people by exploiting their vulnerabilities. There

are two major factors which make consumers natu-

rally vulnerable to fake news (Shu et al., 2017a; Zhou

et al., 2019b): 1) Naive Realism as people tend to be-

lieve that their perceptions of reality are the only ac-

curate views, while others who disagree are regarded

as uninformed or irrational and 2) Conﬁrmation Bias

as people prefer to receive information that conﬁrms

their beliefs.

Moreover, people are more susceptible to certain

kinds of (fake) news due to to the way newsfeed ap-

pears on their homepages in social media thus am-

plifying the psychological challenges to dispelling

fake news. Indeed, people trust fake news owing to

two psychological factors(Shu et al., 2017a): 1) so-

cial credibility, which means people are more likely

to recognize a source as trustworthy if others recog-

nize the source as reliable, mainly when there is not

enough information to assess the truthfulness of the

source; 2) frequency heuristic, which means that peo-

ple may obviously favour information they hear re-

peatedly, although it is fake news.

Our Approach in a Nutshell. Fake news detection

problem can be formalized as a classiﬁcation task thus

requiring features extraction and model construction

sub-taks. The detection phase is a crucial task as it

is devoted to guarantee users to receive authentic in-

formation. We will focus on ﬁnding clues from news

contents.

Our goal is to improve the existing approaches de-

ﬁned so far when fake news is intentionally written

to mislead users by mimicking true news. More in

detail, traditional approaches are based on veriﬁca-

tion by human editors and expert journalists but do

not scale to the volume of news content that is gener-

ated in online social networks. As a matter of fact, the

huge amount of data to be analyzed calls for the devel-

opment of new computational techniques. It is worth

noticing that, such computational techniques, even if

the news is detected as fake, require some sort of ex-

pert veriﬁcation before being blocked. In our frame-

work, we perform an accurate pre-processing of news

data and then we apply several approaches for ana-

lyzing text and multimedia contents. The approach

we discuss in detail in this paper is based on machine

DATA 2020 - 9th International Conference on Data Science, Technology and Applications

152

learning techniques. In this respect, we implemented

several algorithms and we compared them as will be

better explained in the experimental section in order

to ﬁnd out the most suitable one for the fake news

scenario.

2 OUR FAKE NEWS DETECTION

FRAMEWORK

Our framework is based on news ﬂow processing and

data management in a pre-processing block which

performs ﬁltering and aggregation operation over the

news content. Moreover, ﬁltered data are processed

by two independent blocks: the ﬁrst one performs nat-

ural language processing over data while the second

one performs a multimedia analysis.

The overall process we execute for fake news de-

tection is depicted in Figure1.

Figure 1: The overall process at a glance.

In the following we describe each module in more

detail.

Data Ingestion Module. This module take care of

data collection tasks. Data can be highly heteroge-

neous: social network data, multimedia data and news

data. We collect the news text and eventual related

contents and images.

Pre-processing Module. This component is devoted

to the acquisition of the incoming data ﬂow. It per-

forms ﬁltering, data aggregation, data cleaning and

enrichment operations.

NLP Processing Module. It performs the crucial task

of generating a binary classiﬁcation of the news arti-

cles, i.e., whether they are fake or reliable news. It is

split in two submodules. The Machine Learning mod-

ule performs classiﬁcation using an ad-hoc imple-

mented Logistic Regression algorithm (the rationale

for this choice will be explained in the experimental

section) after an extensive process of feature extrac-

tion and selection TF-IDF based in order to reduce

the number of extracted features. The Deep Learning

module classify data using Google Bert algorithm af-

ter a tuning phase on the vocabulary. It also perform

a binary transformation and eventual text padding in

order to better analyze the input data.

Multimedia Processing Module. This module is

tailored for Fake Image Classiﬁcation through Deep

Learning algorithms, using ELA (Error Level Analy-

sis) and CNN.

Due to space limitation, we discuss in the follow-

ing only the details of the deep learning module and

the obtained results.

2.1 The Software Architecture

The software implementation of the framework de-

scribed above is shown in Figure 2.

Figure 2: Our fake news detection framework.

Herein: the data ingestion block is implemented

by using several tools. As an example for Twitter data

we leverage Tweepy

, a Python library to access the

Twitter API. All tweets are downloaded through this

library. Filtering and aggregation is performed us-

ing Apache Kafka

which is able to build real-time

data pipelines and streaming apps. It is scalable,

fault-tolerant and fast thus making our prototype well-

suited for huge amount of data.

The data crawler uses the Newspaper Python li-

brary

whose purpose is extracting and curating arti-

cles. The analytical data archive stores pre-processed

data that are used for issuing queries by traditional

analytical tools. We leverage Apache Cassandra

datastore because it provides high scalability, high

availability, fast writing, fault-tolerance on commod-

ity hardware or cloud infrastructure. The data ana-

lytics block retrieves news contents and news images

from Cassandra DB that are pre-processed by the Ma-

chine Learning module using Scikit Learn library

and by Deep Learning module using Keras library

Image content is processed by the Multimedia Deep

Learning module using Keras library.

In the following we will brieﬂy describe how the

overall process is executed. Requests to the Cassandra

https://www.tweepy.org/

https://kafka.apache.org/

https://newspaper.readthedocs.io/en/latest/

http://cassandra.apache.org/

https://scikit-learn.org/stable/

https://keras.io/

Leveraging Machine Learning for Fake News Detection

153

DB are made through remote access. Each column in

Cassandra refers to a speciﬁc topic and contains all

news belonging to that topic. Among all news, those

having a valid external link value are selected. In this

way, the news content can be easily crawled. As the

link for each news is obtained, a check is performed in

order to verify the current state of the website. If the

website is still running, we perform the article scrap-

ing. The algorithm works by downloading and pars-

ing the news article, then, for each article, title, text,

authors, top image link, news link data are extracted

and saved as a JSON ﬁle in Cassandra DB.

Finally, three independent analysis are then per-

formed by three ad-hoc Python modules we imple-

mented. The ﬁrst two perform text classiﬁcation, and

the last one images classiﬁcation. Concerning the text

analysis, the problem being solved is a binary classi-

ﬁcation one where class 0 refers to reliable news and

class 1 refers to fake ones.

2.1.1 The Machine Learning Module

As mentioned above the goal of the Machine Learn-

ing Module is to produce a binary classiﬁcation on a

text dataset. Thus, a news article will be labelled as

0 if it is recognised as Real, and as 1 if it is recog-

nised as Fake. We devise a supervised approach since

the dataset we worked on is fully labelled. The Ma-

chine Learning implementation has been chosen by

comparing most of the available classiﬁers provided

by the Scikit- Learn library.

It has been developed using a Python 3 kernel in

Jupyter, that is a web-based interactive development

environment for code, and data. It is ﬂexible, ex-

tensible, modular and conﬁgurable to support a wide

range of workﬂows in data science, scientiﬁc com-

puting, and machine learning. We choose Python as it

is interactive, interpreted, modular, dynamic, object-

oriented, portable and extensible thus offering an high

ﬂexibility for our purposes.

More in detail, the following libraries have been

used: i) Scikit- Learn: a simple and efﬁcient tool for

data mining and data analysis; ii) Spacy: an open-

source software library for advanced Natural Lan-

guage Processing, iii) Numpy: a library for Python

programming language that offer support for large,

multi- dimensional arrays and matrices, along with

a large collection of high level mathematical func-

tions to operate on these arrays, iv) Pandas: an

open source, BSD-licensed library providing high-

performance, easy-to-use data structures and data

analysis tools for the Python programming and v)

Matplotlib: a plotting library for the Python program-

ming language and its numerical mathematics exten-

sion NumPy.

In order to choose the most suitable classiﬁca-

tion method, we performed an extensive tuning phase

by comparing several algorithms. The performances

have been evaluated on several test sets by compar-

ing several accuracy measure like Accuracy, Preci-

sion, Recall, F1 measure, Area Under Curve (AUC)

(Flach and Kull, 2015) (reported in Figure 3) and ex-

ecution times (reported in Figure 4). The results are

shown for LIAR datasets but the same trend has been

observed for each data set being analyzed.

As it is easy to see, the best model in terms of

accuracy turns out to be Logistic Regression, so we

decided to perform a parameter optimization for this

algorithm as it exhibits the best results on each efﬁ-

ciency and effectiveness measure. It is worth noticing

that, classiﬁers based on tree construction executes

much slower because of the training step.

We brieﬂy recall here, that logistic regression

instead is a statistical model that leverages the logit

function to model a binary dependent variable [13],

i.e., a linear combination of the observed features:

log

1−p

= β

+ β

˙x

Logistic Regression outputs the probabilities of a

speciﬁc class that are then used for class predictions.

The logistic function exhibits two interesting proper-

ties for our purposes: 1) it has a regular “s” shape; 2)

Its output is bounded between 0 and 1.

Compared with other models, Logistic Regression

offers the following advantages: 1) it is easily inter-

pretable; 2) Model training and prediction steps are

quite fast; 3) Only few parameters has to be tuned

( the regularization parameter); 4) It performs well

even with small datasets; 5) It outputs well-calibrated

predicted probabilities. Nevertheless, there are some

drawbacks as the need of a linear relationship between

the features and the log-odds of the response and it is

not able to automatically learn feature interactions In

order to tune the algorithm we leveraged the function-

alities offered by SciKit.

3 OUR BENCHMARK

In this section we will describe the fake news detec-

tion process and the datasets we used as a benchmark

for our algorithms.

3.1 Dataset Description

Liar Dataset. This dataset includes 12.8K human

labelled short statements from fact-checking website

DATA 2020 - 9th International Conference on Data Science, Technology and Applications

154

Figure 3: Classiﬁer Effectiveness Comparison.

Figure 4: Execution Times comparison.

Politifact.com. Each statement is evaluated by a Poli-

tifact.com editor for its truthfulness. The dataset has

six ﬁne-grained labels: pants-ﬁre, false, barely-true,

half-true, mostly-true, and true. The distribution of la-

bels is relatively well- balanced. For our purposes the

six ﬁne-grained labels of the dataset have been col-

lapsed in a binary classiﬁcation, i.e., label 1 for fake

news and label 0 for reliable ones. This choice has

been made due to binary Fake News Dataset feature.

The dataset is partitioned into three ﬁles: 1) Train-

ing Set: 5770 real news and 4497 fake news; 2) Test

Set: 1382 real news and 1169 fake news; 3) Validation

Set: 1382 real news and 1169 fake news. In Figure5

we show the distribution of real and fake news for the

test dataset.

The three subsets are well balanced so there is no

need to perform oversampling or undersampling. The

corresponding Wordclouds for fake news is reported

in Figure 6. It is easy to see that news are mainly re-

lated to United States. Fake news topics are collected

Figure 5: LIAR Test Dataset in a short.

Figure 6: Liar Fake Wordclouds.

Figure 7: Liar Real Wordclouds.

about Obama, Obamacare, Cicilline, Romney.

On the other side real news topics depicted in Fig-

ure 7 refer to McCain, elections and Obama.

The processed dataset has been uploaded in

Google Drive and, then, loaded in Colab’s Jupyter

as a Pandas Dataframe. It has been added a new

column with the number of words for each row ar-

ticle. By this column it is possible to obtain the fol-

lowing statistical information: count 15389.000000,

mean 17.962311, std 8.569879, min 1.000000, 25%

12.000000, 50% 17.000000, 75% 22.000000, max

66.000000. These statistics show that there are ar-

ticles with only one word in the dataset, so it has

been decided to remove all rows with less than 10

Leveraging Machine Learning for Fake News Detection

155

words as they are considered poorly informative. The

resulting dataset contains 1657 less rows than the

original one. The updated statistics are reported in

what follows: count 13732.000000, mean 19.228663,

std 8.192268, min 10.000000, 25% 14.000000, 50%

18.000000, 75% 23.000000, max 66.000000. Finally,

the average number of words per article is 19.

FakeNewsNet. This dataset has been built by gath-

ering information from two fact-checking websites to

obtain news contents for fake news and real news such

as PolitiFact and GossipCop. In PolitiFact, journalists

and domain experts review the political news and pro-

vide fact-checking evaluation results to claim news

articles as fake or real. Instead, in GossipCop, enter-

tainment stories, from various media outlets, are eval-

uated by a rating score from on the scale of 0 to 10 as

the degree from fake to real. The dataset contains 900

political news and 20k gossip news and has only two

labels: true and false.

This dataset is publicly available by the functions

provided by the FakeNewsNet team and the Twitter

API. As mentioned above, FakeNewsNet can be split

in two subsets: GossipCop and Politifact.com. We

decided to analyse only political news as they pro-

duce worse consequences in real world than gossip

ones. The dataset is well balanced and contains 434

real news and 367 fake news. Most of the news re-

gards the US as it has already been noticed in LIAR.

Fake news topics concern Obama, police, Clinton and

Trump while real news topics refer to Trump, Repub-

licans and Obama. Such as the LIAR dataset, it has

been added a new column and the following statisti-

cal information have been obtained: count 801, mean

1459.217228, std 3141.157565, min 3, 25% 114, 50%

351, 75% 893, max 17377.

The average number of words per articles in Poli-

tifact dataset is 1459, which is far longer than the av-

erage sentence length in Liar Dataset that is 19 words

per articles. Such a statistics conﬁrmed our belief that

it would be better to compare the model performances

on datasets with such different features.

4 EVALUATION

In order to show that the model we implemented out-

performs the results of the current approaches, we

preliminary report in Figure 8 the best results ob-

tained for the other approaches commonly used in lit-

erature for LIAR datasets and in Figure 9 the results

we obtained for the logistic regression algorithm we

implemented that gave us the best results in the fake

news scenario.

We compared the performances on well-

established evaluation measure like: Accuracy,

Precision, Recall, F1 measure, Area Under Curve

(AUC) (Flach and Kull, 2015) and the values re-

ported in the obtained confusion matrices for each

algorithm, i.e., True Positive (TP), False Positive

(FP), True Negative (TN) and False Negative (FN).

Figure 8: Comparison against state of the art approaches on

LIAR datataset.

Figure 9: Our results on LIAR datataset.

We hypothesize that our results are quite better

due to the ﬁne feature selection we performed, a better

pre-processing step and the proper text transformation

and loading.

Figure 10: Our results on Polifact datataset.

In Figure 10 we report the results we obtained on

Polifact dataset.

For the sake of completeness, we report in Figure

11 and Figure 12 the detailed confusion matrices ob-

tained for LIAR and Polifact datasets.

5 CONCLUSION AND FUTURE

WORK

In this paper, we investigated the problem of fake

news detection by machine learning algorithms. We

developed a framework the leverage several algo-

rithms for analyzing real-life datasets and the results

we obtained are quite encouraging. In particular, we

DATA 2020 - 9th International Conference on Data Science, Technology and Applications

156

Figure 11: Confusion Matrix for LIAR dataset.

Figure 12: Confusion Matrix for Polifact dataset.

found that the most accurate results can be obtained

with logistic regression based algorithms. As a future

work, we would like to extend our analysis by better

considering also user proﬁles’ features and some kind

of dynamic analysis of news diffusion mechanism in

our fake news detection model.

REFERENCES

Agrawal et al., D. (2012). Challenges and opportunities

with big data. A community white paper developed

by leading researchers across the United States. Tech-

nical report, Purdue University.

Allcott, H. and Gentzkow, M. (2017). Social media and

fake news in the 2016 election. Working Paper 23089,

National Bureau of Economic Research.

Boselli, R., Cesarini, M., Mercorio, F., and Mezzanzan-

ica, M. (2018). Classifying online job advertisements

through machine learning. Future Generation Com-

puter Systems, 86:319 – 328.

Cassavia, N., Masciari, E., Pulice, C., and Sacc

a, D. (2017).

Discovering user behavioral features to enhance infor-

mation search on big data. TiiS, 7(2):7:1–7:33.

Culpepper, J. S., Moffat, A., Bennett, P. N., and Lerman, K.,

editors (2019). Proceedings of the Twelfth ACM Inter-

national Conference on Web Search and Data Min-

ing, WSDM 2019, Melbourne, VIC, Australia, Febru-

ary 11-15, 2019. ACM.

Flach, P. A. and Kull, M. (2015). Precision-recall-gain

curves: PR analysis done right. In Cortes, C.,

Lawrence, N. D., Lee, D. D., Sugiyama, M., and Gar-

nett, R., editors, Advances in Neural Information Pro-

cessing Systems 28: Annual Conference on Neural In-

formation Processing Systems 2015, December 7-12,

2015, Montreal, Quebec, Canada, pages 838–846.

Guo, C., Cao, J., Zhang, X., Shu, K., and Yu, M. (2019).

Exploiting emotions for fake news detection on social

media. CoRR, abs/1903.01728.

Masciari, E. (2012). SMART: stream monitoring enterprise

activities by RFID tags. Inf. Sci., 195:25–44.

Masood, R. and Aker, A. (2018). The fake news challenge:

Stance detection using traditional machine learning

approaches.

Mezzanzanica, M., Boselli, R., Cesarini, M., and Mercorio,

F. (2015). A model-based approach for developing

data cleansing solutions. Journal of Data and Infor-

mation Quality (JDIQ), 5(4):1–28.

Potthast, M., Kiesel, J., Reinartz, K., Bevendorff, J., and

Stein, B. (2017). A stylometric inquiry into hyperpar-

tisan and fake news. CoRR, abs/1702.05638.

Shabani, S. and Sokhn, M. (2018). Hybrid machine-crowd

approach for fake news detection. In 2018 IEEE 4th

International Conference on Collaboration and Inter-

net Computing (CIC), pages 299–306. IEEE.

Shu, K., Sliva, A., Wang, S., Tang, J., and Liu, H. (2017a).

Fake news detection on social media: A data mining

perspective. CoRR, abs/1708.01967.

Shu, K., Sliva, A., Wang, S., Tang, J., and Liu, H. (2017b).

Fake news detection on social media: A data mining

perspective. CoRR, abs/1708.01967.

Shu, K., Wang, S., and Liu, H. (2019). Beyond news con-

tents: The role of social context for fake news detec-

tion. In (Culpepper et al., 2019), pages 312–320.

Vosoughi, S., Roy, D., and Aral, S. (2018). The spread of

true and false news online. Science, 359(6380):1146–

1151.

Zhou, X., Zafarani, R., Shu, K., and Liu, H. (2019a). Fake

news: Fundamental theories, detection strategies and

challenges. In (Culpepper et al., 2019), pages 836–

837.

Zhou, Z., Guan, H., Bhat, M. M., and Hsu, J. (2019b). Fake

news detection via nlp is vulnerable to adversarial at-

tacks. arXiv preprint arXiv:1901.09657.

Leveraging Machine Learning for Fake News Detection

157