Fake News Detector 2

A INTERNSHIP REPORT ON
“FAKE NEWS PREDICTION USING
MACHINE LEARNING”

Internship report submitted in partial fulfilment of the

requirements for the award of degree of

MASTER OF COMPUTER APPLICATIONS

Accredited by National Board of Accreditation
Submitted by
VEERESH A C
1DA22MC051
Under the Guidance of
Mrs. Anitha J
Associate Professor,

Department of Master of Computer Application,

Dr. AMBEDKAR INSTITUTE OF TECHNOLOGY.

Dr. AMBEDKAR INSTITUTE OF TECHNOLOGY, (AN

AUTONOMOUS INSTITUTION, AFFILIATED TO VTU,
BELAGAVAI)
BDA Outer Ring Road, Mallathahally, Bangalore-560056
 Fake News Detector 3

Dr. AMBEDKAR INSTITUTE OF TECHNOLOGY,

(AN AUTONOMOUS INSTITUTION, AFFILIATED TO VTU,
BELAGAVAI)
BDA Outer Ring Road, Mallathahally, Bangalore-560056

MASTER OF COMPUTER APPLICATIONS

Accredited by National Board of Accreditation

CERTIFICATE

This is to Certify that VEERESH A C bearing 1DA22MC051 has

completed his third semester Internship entitled “FAKE NEWS
PREDICTION USING MACHINE LEARNING” as a partial
fulfilment for the award of Master of Computer Applications degree,
during the academic year 2023-24 under supervision.

Signature of Guide
Mrs. Anitha J
Associate Professor,
Department of Master of Computer Application,
Dr. AMBEDKAR INSTITUTE OF TECHNOLOGY

Head of the Department Principal

 Fake News Detector 4
DECLARATION

VEERESH A C student of 3rd sem MCA. Dr. AMBEDKAR INSTITUTE

OF TECHNOLOGY, bearing USN 1DA22MC051 hereby declare that the
Internship entitled on a “FAKE NEWS PREDICTION USING
MACHINE LEARNING” has been carried out by me under the
supervision of Guide Mrs. Anitha J Associate Professor, and DR.
CHANDRAKANTH G PUJARI, Professor & Head submitted in partial
fulfilment of the requirements for the award of the Degree of Master of
Computer Applications during the academic year 2024. This report has
not been submitted to any other Organization/University for any award
of degree or certificate.

Signature of Student
VEERESH A C
(1DA22MC051)
 Fake News Detector 5

ACKNOWLEDGEMENT

I would like to thank DR. M. MEENAKSHI, Principle, Dr. AIT, who

has always been great source of inspiration and for permitting me to carry
out the internship.

I specially thank to DR. CHANDRAKANTH G PUJARI, Professor and

Head, department of MCA, for his kind cooperation.

I extend my special thanks to Mrs. Anitha J Associate Professor,

department of MCA, he has been constant source of inspiration in
completing the project.

I likewise thank all the faculty individuals and my friends for the help and
consolation I might want to thank my parents and our family individuals who
upheld and helped me in finishing the Internship.

Place: VEERESH A
Date: (1DA22MC051)
Fake News Detector 6

Abstract

With the recent social media boom, the spread of fake news has become a great

concern for everybody. It has been used to manipulate public opinions, influence

the election - most notably the US Presidential Election of 2016, incite hatred and

riots like the genocide of the Rohingya population. A 2018 MIT study found that

fake news spreads six times faster on Twitter than real news. The credibility and

trust in the news media are at an all-time low. It is becoming increasingly difficult

to determine which news is real and which is fake. Various machine learning

methods have been used to separate real news from fake ones. In this study, we

tried to accomplish that using Passive Aggressive Classifier, LSTM and natural

language processing. There are lots of machine learning models but these two

have shown better progress.

Now there is some confusion present in the authenticity of the correctness. But it

definitely opens the window for further research. There are some of the aspects that

has to be kept in mind considering the fact that fake news detection is not only a

simple web interface but also a quite complex thing that includes a lot of backend

work.
Fake News Detector 7

Table of Content

Declaration of Authorship 2
Acknowledgement 3
Abstract 4
Table of Content 5
Introduction 6
2. Problem Statement 7
3. Motivation 8
4. Background Study 11
5. Feasibility Study 13
6. Methodology 14
6.1 The Dataset 14
6.2 The Machine Learning Model 15
6.3 The Web Interface 18
6.4 Common Platform: Flask 18
7. Implementation 19
7.1 The Interface 19
7.2 The ML Model 21
7.3 Flask Code 34
7.4 Web Interface 37
8. Key Insights 42
9. Conclusion 43
10. Future Work 44
11. References 45
Fake News Detector 8

1. Introduction

Fake news is untrue information presented as news. It often has the aim of

damaging the reputation of a person or entity or making money through advertising

revenue. Once common in print, the prevalence of fake news has increased with

the rise of social media, especially the Facebook News Feed. During the 2016 US

presidential election, various kinds of fake news about the candidates widely

spread in the online social networks, which may have a significant effect on the

election results. According to a post-election statistical report, online social

networks account for more than 41.8% of the fake news data traffic in the election,

which is much greater than the data traffic shares of both traditional TV/radio/print

medium and online search engines respectively. Fake news detection is becoming

increasingly difficult because people who have ill intentions are writing the fake

pieces so convincingly that it is difficult to separate from real news. What we have

done is a simplistic approach that looks at the news headlines and tries to predict

whether they may be fake or not.

Fake news can be intimidating as they attract more audience than normal. People

use them because this can be a very good marketing strategy. But the money

earned might not live upto fact that it can harm people.
Fake News Detector 9

2. Problem Statement

In this day and age, it is extremely difficult to decide whether the news we come

across is real or not. There are very few options to check the authenticity and all of

them are sophisticated and not accessible to the average person. There is an acute

need for a web-based fact-checking platform that harnesses the power of Machine

Learning to provide us with that opportunity.

Fake News Detector 10

3. Motivation

Social media facilitates the creation and sharing of information that uses computer-

mediated technologies. This media changed the way groups of people interact and

communicate. It allows low cost, simple access and fast dissemination of

information to them. The majority of people search and consume news from social

media rather than traditional news organizations these days. On one side, where

social media have become a powerful source of information and bringing people

together, on the other side it also 1 put a negative impact on society. Look at some

examples herewith; Facebook Inc’s popular messaging service, WhatsApp became

a political battle-platform in Brazil’s election. False rumours, manipulated photos,

de-contextualized videos, and audio jokes were used for campaigning. These kinds

of stuff went viral on the digital platform without monitoring their origin or reach.

A nationwide block on major social media and messaging sites including Facebook

and Instagram was done in Sri Lanka after multiple terrorist attacks in the year

2019. The government claimed that “false news reports” were circulating online.

This is evident in the challenges the world's most powerful tech companies face in

reducing the spread of misinformation. Such examples show that Social Media

enables the widespread use of “fake news” as well. The news

Fake News Detector 11

disseminated on social media platforms may be of low quality carrying misleading

information intentionally. This sacrifices the credibility of the information.

Millions of news articles are being circulated every day on the Internet – how one

can trust which is real and which is fake? Thus incredible or fake news is one of

the biggest challenges in our digitally connected world. Fake news detection on

social media has recently become an emerging research domain. The domain

focuses on dealing with the sensitive issue of preventing the spread of fake news

on social media. Fake news identification on social media faces several challenges.

Firstly, it is difficult to collect fake news data. Furthermore, it is difficult to label

fake news manually. Since they are intentionally written to mislead readers, it is

difficult to detect them simply based on news content. Furthermore, Facebook,

Whatsapp, and Twitter are closed messaging apps. The misinformation

disseminated by trusted news outlets or their friends and family is therefore

difficult to be considered as fake. It is not easy to verify the credibility of newly

emerging and time-bound news as they are not sufficient to train the application

dataset. Significant approaches to differentiate credible users, extract useful news

features and develop authentic information dissemination systems are some useful

domains of research and need further investigations. If we can’t control the spread

of fake news, the trust in the system will collapse. There will be widespread
Fake News Detector 12

distrust among people. There will be nothing left that can be objectively used. It

means the destruction of political and social coherence. We wanted to build some

sort of web-based system that can fight this nightmare scenario. And we made

some significant progress towards that goal.

Fake News Detector 13

4. Background Study

From an NLP perspective, researchers have studied numerous aspects of the

credibility of online information. For example, [1] applied the time-sensitive

supervised approach by relying on tweet content to address the credibility of a

tweet in different situations. [2] used LSTM in a similar problem of early rumour

detection. In another work, [3] aimed at detecting the stance of tweets and

determining the veracity of the given rumour with convolution neural networks. A

submission [4] to the SemEval 2016 Twitter Stance Detection task focuses on

creating a bag-of-words autoencoder and training it over the tokenized tweets.

Another team, [5], combined multiple models in an ensemble providing a 50/50

weighted average between a deep convolutional neural network and a gradient-

boosted decision tree. Though this work seems to be similar to our work, the

difference lies in the construction of an ensemble of classifiers. In a similar

attempt, a team [6] concatenated various features vectors and passed them through

an NLP model. Passive Aggressive algorithm is a margin-based online learning

algorithm for binary classification. It is also an algorithm of a soft margin-based

method and robust to noise. It can be used in fake news detection [16] Term

Frequency-Inverse Document Frequency is also a method used to represent text in

Fake News Detector 14

a format that can be easily processed by machine learning algorithms. It is a

numerical statistic that shows how important a word is to news in a news dataset.

The importance of a word is proportional to the number of times the word appears

in the news (fake and real) but inversely proportional to the number of times the

word appears in the news dataset (fake or real) [15]

Fake News Detector 15

5. Feasibility Study

Passive-aggressive classifier, logistic regression, LSTM can be used in fake news

detection. Bi-directional LSTM was used in [7] to detect fake news. It had

reasonably good accuracy but if the news was a bit more sophisticated, it would be

difficult to achieve good accuracy. Because this model picks up the

sensational/clickbaity words as part of fake news. For example, if a news title says,

‘Donald Trump is the greatest president ever, the model will pick it up as fake

news with reasonable accuracy. If the title is more nuanced and written in a

sophisticated way, it’d be difficult to do so. We believe that our LSTM model is not

enough by itself to detect fake news. That’s why we included passive aggressive

classifier with it and when we compared passive news with reputable news

sources, but the scope of the work is so vast that we couldn’t do it with the

resources available to us. Our model can act as a first step in detecting fake news.

But more work is needed to call the model reliable enough.

Fake News Detector 16

6. Methodology

6.1 The Dataset

Figure 1 : Dataset

The dataset is simple. It contains the titles of the news, the body text and a label

field, which, if the news is authentic, shows REAL and if inauthentic, shows

FAKE.

There are 3 main segments of the methodology :

◦ The core Machine Learning model.

Fake News Detector 17

◦ The web interface.

◦ The common platform that brings the model and the interface together.

6.2 The Machine Learning Model

There are two parts to the ML Model building. Machine Learning is a part of our

life that can help us in predicting. We are using two types of model in this case. For

the first part, we used passive-aggressive classifiers. And the steps include:

1. Data Loading: We are loading a CSV file for the data sorting and training-

testing part of the model. The CSV file is turned into an array for easier

work purpose.

2. Vectorization: Vectorization is needed for determining the frequency of the

words present in a passage. This is needed to determine which words are

used often.

3. Classifier: Passive-aggressive algorithms are a family of great learning

algorithms. They are similar to Perceptron because it does not require a

reading scale. However, unlike Perceptron, they include parameter

correction. Passive is used when the prediction is correct and there is no

Fake News Detector 18

change in the model. But if there is any kind of change in the model, that is

if the prediction is not correct then the aggressive part is called, which

changes the model accordingly. The aggressive part of the model changes

the model according to its wish on the backend.

Figure 2 : Passive-aggressive model

4. Model Building: The model is built through the train and test of the dataset,

by ensuring that the training is done for 80% of the dataset and testing is

done in the rest of the 20% of the dataset.

In the second part, we used is LSTM. Here are the steps :

1. Loading the data: For this step, it is the same as the passive-aggressive

one.
Fake News Detector 19

2. Scanning and parsing. Data is loaded from a CSV file. This consists of

the body of selected news articles. It then contains a label field that indicates

whether the news is real or fake. In this code block, we scan the CSV and

clean the titles to filter out stop words and punctuation.

3. Tokenization. The tokenizer is used to assign indices to words, and filter

out infrequent words. This allows us to generate sequences for our training

and testing data.

4. Embedding matrix: Apply the embedding matrix. An embedding matrix

is used to extract the semantic information from the words in each title.

5. Model Building: Building the model and finding out the accuracy via

confusion matrix. The model is created using an Embedding layer, LSTM,

Dropout, and Dense layers. We are going to run the data on 20 epochs.

We observed that the LSTM model is vastly inaccurate in predicting the

authenticity of the news. So we decided to show the output by running it

through the Passive-aggressive classifier model.

Fake News Detector 20

6.3 The Web Interface

This was the simplest part.

1. HTML for building the basic skeleton: HTML makes the structure of the

web application and also there are some of the functions that can be

achieved best with HTML only.

2. CSS for design: The CSS part is for designing only. Because it will give a

more beautiful aspect to the website.

6.4 Common Platform: Flask

This acts as a common platform and takes the input with the pickle module and

passes it to the machine learning model afterwards the prediction is shown on the

screen with the HTML and CSS website.

1. Building functions for taking input.

2. Passing input values through the ML model.

3. Using the Pickle module for serializing and de-serializing the dataset.

4. Providing output.
Fake News Detector 21

7. Implementation

7.1 The Interface

This is what you see when you go to the web interface. You are supposed to copy

the news and paste it into the input box.

Figure 3.1 : The Interface

When you paste the news on the input box and click ‘Predict’ the model will give

you the result. If the news seems authentic, the output will be ‘Looking Real

News’. Otherwise, it will show ‘Looking Fake News’. That’s how you can detect

fake or real news via the interface.

Fake News Detector 22

Figure 3.2 : The Interface

Fake News Detector 23

7.2 The ML Model

The code for the ML model building is as follows:

TF-IDF stands for Term Frequency-Inverse Document Frequency. Term frequency

is basically a ratio of the number of times a particular word appears with respect to

the total number of word. And Inverse Document Frequency is basically the weight

of a rare word.

from sklearn.feature_extraction.text import

TfidfVectorizer

text = ['This is the final project of Mashiat Nahreen,

Lutfor Rafe and Rabiul Alam Abir', 'This is the final

project of our undergrad.' ]

vectorization = TfidfVectorizer()

vectorization.fit(text)

print(vectorization.idf_)

print(vectorization.vocabulary_)

Words that are present in every data will have very low IDF value and using that

we will highlight the maximum IDF values.

example = text[0]
Fake News Detector 24

example

example = vectorization.transform([example])

print(example.toarray())

The zeros represent there are no words in that postion.

IMPLEMENTING PASSIVE AGGRESSIVE CLASSIFIER

Passive is used when the prediction is correct and there is no change in the model.

But if there is any kind of change in the model that is if the prediction is not correct

then aggressive part is called, which changes the model accordingly.

import os

os.chdir("D:\Books\Fake_News_Detection-master")

OS module is used for the Python program to interact with the operating system

import pandas as pd

dataset = pd.read_csv('news.csv')

dataset.head()

x = dataset['text']

y = dataset['label']

from sklearn.model_selection import train_test_split

Fake News Detector 25

from sklearn.feature_extraction.text import

TfidfVectorizer

from sklearn.linear_model import

PassiveAggressiveClassifier

from sklearn.metrics import accuracy_score,

confusion_matrix

x_train,x_test,y_train,y_test =

train_test_split(x,y,test_size=0.2,random_state=0)

y_train

y_train

vectorization =

TfidfVectorizer(stop_words='english',max_df=0.7)

xv_train = vectorization.fit_transform(x_train)

xv_test = vectorization.transform(x_test)

max_df refers to the percentage of the repetition of the word. 0.7 means 70% of the

time the word is repeated.

classifier = PassiveAggressiveClassifier(max_iter=50)

classifier.fit(xv_train,y_train)

y_pred = classifier.predict(xv_test)
Fake News Detector 26

score = accuracy_score(y_test,y_pred)

print(f'Accuracy: {round(score*100,2)}%')

cf = confusion_matrix(y_test,y_pred,

labels=['FAKE','REAL'])

print(cf)

def fake_news_det(news):

input_data = [news]

vectorized_input_data =

vectorization.transform(input_data)

prediction =

classifier.predict(vectorized_input_data)

print(prediction)

fake_news_det('U.S. Secretary of State John F. Kerry

said Monday that he will stop in Paris later this

week, amid criticism that no top American officials

attended Sunday’s unity march against terrorism.')

fake_news_det("""Go to Article

President Barack Obama has been campaigning hard for

the woman who is supposedly going to extend his legacy

Fake News Detector 27

four more years. The only problem with stumping for

Hillary Clinton, however, is she’s not exactly a

candidate easy to get too enthused about. """)

import pickle

pickle.dump(classifier,open('model.pkl', 'wb'))

pickle is used for serializing and deserializing any

data that is inputted in Python.

loaded_model = pickle.load(open('model.pkl', 'rb'))

def fake_news_det1(news):

input_data = [news]

vectorized_input_data =

vectorization.transform(input_data)

prediction =

classifier.predict(vectorized_input_data)

print(prediction)

fake_news_det1("""U.S. Secretary of State John F.

Kerry said Monday that he will stop in Paris later

this week, amid criticism that no top American

Fake News Detector 28

officials attended Sunday’s unity march against

terrorism.""")

fake_news_det('''U.S. Secretary of State John F. Kerry

said Monday that he will stop in Paris later this

week, amid criticism that no top American officials

attended Sunday’s unity march against terrorism.''')

In this project, titles of news articles found on the internet is used to determine

whether a news is fake or real. We are using LSTM to help classify them into either

real or fake category.

import numpy as np

import pandas as pd

import json as j

import urllib

import gzip

import nltk

nltk.download('stopwords')

from nltk.stem import PorterStemmer

from sklearn.model_selection import train_test_split

!pip install gensim

Fake News Detector 29

from gensim.models import KeyedVectors

from nltk.corpus import stopwords

from keras.models import Model

from keras.callbacks import EarlyStopping,

ModelCheckpoint

from keras.layers import Dense, Input, LSTM,

Embedding, Dropout, Activation

from keras.layers.merge import concatenate

from keras.layers.normalization import

BatchNormalization

from keras.preprocessing import sequence

from keras.preprocessing.text import Tokenizer

from keras.preprocessing.sequence import pad_sequences

Data scanning and parsing : Data is loaded from a csv file fake_or_real_news.csv.

This consists of the title and text of a select group of news articles. It then contains

a label field which indicates whether the news is real or fake. In this code block,

we scan the csv and clean the titles to filter out stop words and punctuation.

import re

import string
Fake News Detector 30

from sklearn.feature_extraction.text import

CountVectorizer

def clean_text(text):

text = str(text)

text = text.split()

words = []

for word in text:

exclude = set(string.punctuation)

word = ''.join(ch for ch in word if ch not in

exclude)

if word in stops:

continue

try:

words.append(ps.stem(word))

except UnicodeDecodeError:

words.append(word)

text = " ".join(words)

return text.lower()
Fake News Detector 31

stops = set(stopwords.words("english"))

ps = PorterStemmer()

f = pd.read_csv('news.csv')

f.label = f.label.map(dict(REAL=1, FAKE=0))

We take the news titles and divide the train and test set. We also clean the text.

f = f[1:100]

X_train, X_test, y_train, y_test =

train_test_split(f['title'], f.label, test_size=0.2)

X_cleaned_train = [clean_text(x) for x in X_train]

X_cleaned_test = [clean_text(x) for x in X_test]

X_cleaned_train[0]

Tokenizer : Tokenizer is used to assign indices to words, and filter out infrequent

words. This allows us to generate sequences for our training and testing data.

import tokenize

from keras.preprocessing.text import Tokenizer

MAX_NB_WORDS = 20000

tokenizer = Tokenizer(num_words=MAX_NB_WORDS)
Fake News Detector 32

tokenizer.fit_on_texts(X_cleaned_train +

X_cleaned_test)

print('Finished Building Tokenizer')

train_sequence =

tokenizer.texts_to_sequences(X_cleaned_train)

print('Finished Tokenizing Training')

test_sequence =

tokenizer.texts_to_sequences(X_cleaned_test)

print('Finished Tokenizing Training')

Embedding Matrix : Embedding matrix is used to extract the semantic information

from the words in each title.

from gensim.models import KeyedVectors

from gensim.models import Word2Vec

EMBEDDING_FILE =

'https://s3.amazonaws.com/dl4j-distribution/GoogleNews

-vectors-negative300.bin.gz'

Word2Vec =

KeyedVectors.load_word2vec_format(EMBEDDING_FILE,

binary=True)
Fake News Detector 33

word_index = tokenizer.word_index

print('Found %s unique tokens' % len(word_index))

nb_words = min(20000, len(word_index))

embedding_matrix = np.zeros((nb_words, 300))

for word, i in word_index.items():

try:

embedding_vector = word2vec.word_vec(word)

if embedding_vector is not None and i < 7000:

embedding_matrix[i] = embedding_vector

except (KeyError, IndexError) as e:

continue

Building the Model : The model is created using an Embedding layer, LSTM,

Dropout, and Dense layers.We are going to run the data on 20 epochs.

from keras.models import Sequential

from keras.layers import Dense, LSTM, Dropout, Conv1D,

MaxPooling1D

from keras.layers.embeddings import Embedding

from keras.preprocessing import sequence

from keras.preprocessing.sequence import pad_sequences

Fake News Detector 34

kVECTORLEN = 50

model = Sequential()

model.add(Embedding(5000, 500, input_length=50))

model.add(Dropout(0.4))

model.add(Dense(1, activation='relu'))

model.compile(loss='binary_crossentropy',

optimizer='adam', metrics=['accuracy'])

print(model.summary())

train_sequence =

sequence.pad_sequences(train_sequence, maxlen=50)

test_sequence = sequence.pad_sequences(test_sequence,

maxlen=50)

history = model.fit(train_sequence, y_train,

validation_data=(test_sequence, y_test), epochs=20,

batch_size=64)

Calculating the accuracy.

scores = model.evaluate(test_sequence, y_test,

verbose=0)
Fake News Detector 35

accuracy = (scores[1]*100)

print("Accuracy: {:.2f}%".format(scores[1]*100))

Analyzing the Data: The graphs below demonstrate the change in accuracy and

loss for the training data as well as the validation data.

import matplotlib.pyplot as plt

plt.plot(history.history['accuracy'])

plt.plot(history.history['val_accuracy'])

plt.title('model accuracy')

plt.ylabel('accuracy')

plt.xlabel('epoch')

plt.legend(['train', 'validation'], loc='upper left')

plt.show()

plt.plot(history.history['loss'])

plt.plot(history.history['val_loss'])

plt.title('model loss')

plt.ylabel('loss')

plt.xlabel('epoch')

plt.legend(['train', 'test'], loc='upper left')

plt.show()
Fake News Detector 36
Fake News Detector 37

7.3 Flask Code

from flask import Flask, render_template, request

from sklearn.feature_extraction.text import

TfidfVectorizer

from sklearn.linear_model import

PassiveAggressiveClassifier

import pickle

import pandas as pd

from sklearn.model_selection import train_test_split

app = Flask(__name__)

vectorization = TfidfVectorizer(stop_words='english',

max_df=0.7)

loaded_model = pickle.load(open('model.pkl', 'rb'))

dataset = pd.read_csv('news.csv')

x = dataset['text']

y = dataset['label']

x_train, x_test, y_train, y_test = train_test_split(x,

y, test_size=0.2, random_state=0)
Fake News Detector 38

def fake_news_det(news):

xv_train = vectorization.fit_transform(x_train)

xv_test = vectorization.transform(x_test)

input_data = [news]

vectorized_input_data =

vectorization.transform(input_data)

prediction =

loaded_model.predict(vectorized_input_data)

return prediction

@app.route('/')

def home():

return render_template('index.html')

@app.route('/predict', methods=['POST'])

def predict():

if request.method == 'POST':

message = request.form['message']

pred = fake_news_det(message)

print(pred)
Fake News Detector 39

return render_template('index.html',

prediction=pred)

else:

return render_template('index.html',

prediction="Something went wrong")

if __name__ == '__main__':

app.run(debug=True)
Fake News Detector 40

7.4 Web Interface

<!DOCTYPE html>

<html>

<head>

<meta charset="UTF-8">

<title>Fake News Detection System</title>

<link

href='https://fonts.googleapis.com/css?family=Pacifico

' rel='stylesheet' type='text/css'>

<link

href='https://fonts.googleapis.com/css?family=Arimo'

rel='stylesheet' type='text/css'>

<link

href='https://fonts.googleapis.com/css?family=Hind:300

' rel='stylesheet' type='text/css'>

<link

href='https://fonts.googleapis.com/css?family=Open+San

s+Condensed:300' rel='stylesheet' type='text/css'>

Fake News Detector 41

<meta name="viewport" content="width=device-width,

initial-scale=1">

<style>

input[type=text], select, textarea {

width: 50%;

padding: 10px;

border: 3px solid #ccc;

border-radius: 1px;

box-sizing: border-box;

margin-top: 6px;

margin-bottom: 16px;

resize: horizontal;

button {

background-color: #4CAF50;

color: white;

padding: 14px 20px;

margin: 8px 0;

border: none;
Fake News Detector 42

cursor: pointer;

width: 50%;

button:hover {

opacity: 0.8;

h1 {

text-align: center;

p {

text-align: center;

div {

text-align: center;

body {
Fake News Detector 43

background: rgba(0, 128, 0, 0.3) /* Green

background with 30% opacity */

</style>

</head>

<body>

<p style="padding: 0 10em 10em 0">

<div class="login">

<h1 style="text-align:center;">Fake News Detector

</br> By Lutfor Rafe(154429) </br> Rabiul Alam

Abir(160041026) </br> Mashiat

Nahreen(160041028) </h1>

<form action="{{ url_for('predict')}}"

method="POST">

<textarea name="message" rows="6" cols="20"

required="required" style="font-size:

18pt"></textarea>

<br> </br>
Fake News Detector 44

<button type="submit" class="btn btn-primary

btn-block btn-large">Predict</button>

<div class="results">

{% if prediction == ['FAKE']%}
<h2 style="color:red;">Looking
Spam⚠News📰

</h2>

{% elif prediction == ['REAL']%}

<h2 style="color:green;"><b>Looking Real

News📰</b></h2>

{% endif %}

</div>

</form>

</div>

</p>

</body>

</html>
Fake News Detector 45

8. Key Insights

The passive aggressive model produces 93% accuracy. When we input the news

text on the interface, it correctly identifies the news most of the time. We tested this

by using news from The Onion. The Onion is a satire ‘news’ portal that posts fake

funny news. When we pasted some of the news from the site on our web interface,

those were correctly identified as fake. But when we wanted to test the news from

BBC or New York Times, those were correctly identified as real. But the accuracy

of the LSTM model was much lower, so we went with the Passive Aggressive

model to produce output on the interface.

Fake News Detector 46

9. Conclusion

Our project can ring the initial alert for fake news. The model produces worse

results if the article is written cleverly, without any sensationalization. This is a

very complex problem but we tried to address it as much as we could. We believe

the interface provides an easier way for the average person to check the

authenticity of a news. Projects like this one with more advanced features should

be integrated on social media to prevent the spread of fake news.

Fake News Detector 47

10. Future Work

There are many future improvement aspects of this project. Introducing a cross

checking feature on the machine learning model so it compares the news inputs

with the reputable news sources is one way to go. It has to be online and done in

real time, which will be very challenging. Improving the model accuracy using

bigger and better datasets, integrating different machine learning algorithms is also

something we hope to do in the future.

Fake News Detector 48

11. References

[1] C. Castillo, M. Mendoza, and B. Poblete. Predicting information credibility in

time-sensitive social media. Internet Research, 23(5):560–588, 2013.

[2] T. Chen, L. Wu, X. Li, J. Zhang, H. Yin, and Y. Wang. Call attention to

rumours: Deep attention-based recurrent neural networks for early rumour

detection. arXiv preprint arXiv:1704.05973, 2017.

[3] Y.-C. Chen, Z.-Y. Liu, and H.-Y. Kao. Ikm at several-2017 task 8:

Convolutional neural networks for stance detection and rumour verification.

Proceedings of SemEval. ACL, 2017.

[4] I. Augenstein, A. Vlachos, and K. Bontcheva. Usfd at semeval-2016 task 6:

Any-target stance detection on Twitter with autoencoders. In SemEval@NAACL-

HLT, pages 389–393, 2016.

[5] S. B. Yuxi Pan, Doug Sibley. Talos. http://blog.talosintelligence. com/2017/06/,

2017.

[6] B. S. Andreas Hanselowski, Avinesh PVS and F. Caspelherr. Team athene on

the fake news challenge. 2017.

Fake News Detector 49

[7] Bahad, P., Saxena, P. and Kamal, R., 2019. Fake News Detection using Bi-

directional LSTM-Recurrent Neural Network. Procedia Computer Science, 165,

pp.74-82.

[8] EANN: Event Adversarial Neural Networks for Multi-Modal

[9] Fake News Detection on Social Media: A Data Mining Perspective Kai Shuy,

Amy Slivaz, Suhang Wangy, Jiliang Tang \, and Huan Liuy

[10] CSI: A Hybrid Deep Model for Fake News DetectionIdentifying the signs of

fraudulent accounts using data mining techniques Shing-Han Li a,□, David C. Yen

b,1, Wen-Hui Luc,2, Chiang Wanga,2

[11] Automatic Deception Detection: Methods for Finding Fake News. Niall J.

Conroy, Victoria L. Rubin, and Yimin Chen

[15] J. D'Souza, "An Introduction to Bag-of-Words in NLP," 03 04 2018. [Online].

Available:https://medium.com/greyatom/an-introduction-tobag-of-words-in-nlp-ac

967d43b428.

[16] G. Bonaccorso, "Artificial Intelligence – Machine Learning – Data Science,"

10 06 2017.

[Online].Available:https://www.bonaccorso.eu/2017/10/06/mlalgorithms-addendu

m-passive-aggressivealgorithms/