This activity represented the final assignment for the Data Science and Big Data Analysis course within the domain of Natural Language Processing (NLP) at Pavia University. The dataset utilized for this project is accessible through the following link: https://zenodo.org/record/4561253. Professor Antonino Nocera provided guidance throughout the assignment. The collaborative efforts behind the project involve Arnold Fonkou, Vignesh Kumar Kembu, Ashina Nurkoo, and Seyedkourosh Sajjadi.

1. Fake News and Their Detection
Data Science and Big Data Analysis
Professor: Antonino Nocera
Team Name: 4V’s
Group members:
Arnold Fonkou
Vignesh Kumar Kembu
Ashina Nurkoo
Seyedkourosh Sajjadi

2. WELFake
Fake News Detection (WELFake) dataset
of 72,134 news articles with 35,028 real
and 37,106 fake news.
This dataset is a part of an ongoing
research on "Fake News Prediction on
Social Media Website" as a doctoral
degree program of Mr. Pawan Kumar
Verma and is partially supported by the
ARTICONF project funded by the
European Union’s Horizon 2020 research
and innovation program.
Columns:
- Serial number (starting from 0)
- Title (about the text news heading)
Text (about the news content)
- Label (0 = fake and 1 = real)

3. Data Stream Ingestion Hadoop MapReduce
MongoDB Sandbox
PySpark HDFS
Analysis
Architecture
PySpark

4. Ingestion
From CSV to JSON
Data Conversion
we have converted the
file into JSON to be
closer to reality.
Reading Data
Using PySpark
We used the
DATAFRAME client of
SPARK to read our big
data.
Saving to Hadoop
Write into Hadoop
We read from the data
frame and then we write
it to Hadoop.

5. Reading Section
Import findspark
findspark.init()
import pyspark
from pyspark.sql import *
spark = SparkSession.builder
.master("local[1]")
.appName("PySpark Read JSON")
.getOrCreate()
# Reading multiline json file
multiline_dataframe = spark.read.option("multiline","true")
.json("project_data_sample.json")
multiline_dataframe.head()
Saving Section
multiline_dataframe.write.save('/usr/local/hadoop/user3/dsba1.json',
format='json')
And the data is shown as below:
sqlContext = SQLContext(spark)
df = sqlContext.read.format('json').load('/usr/local/hadoop/user3/dsba1.json')
df.show()

6. Hadoop
Hadoop
Component
HDFS
Component
MapReduce

7. Mapper (BoW Creation)
Read Lines
Input Data
The data is given as input
lines to the mapper.
Extract Text
Title and Text Extraction
After reading each line as
a JSON object, we
extract the title and the
text related to that piece
of news from it.
Tokenize
Word Extraction
We perform some data
cleaning and then we
extract every single word
from it.

8. Text Cleaning
import sys
import re
import json
def clean_text(text):
text =
re.sub(r'http[s]?://(?:[a-zA-Z]|[0-9]|[$-_@.&+]|[!*(),]|(
?:%[0-9a-fA-F][0-9a-fA-F]))+', '', text)
text = re.sub(r'[^a-zA-Zs]+', '', text)
return text
Tokenizing
def tokenize(text):
if not isinstance(text, str):
text = str(text)
text = clean_text(text)
text = str.lower(text)
return text.split()
Execution
for line in sys.stdin:
line = line.strip()
Try:
json_obj = json.loads(line)
except:
continue
title = json_obj.get("title", "")
text = json_obj.get("text", "")
title_words = tokenize(title)
text_words = tokenize(text)
for word in title_words + text_words:
print(f"{word}t1")

9. Reducer
Read Lines
Input Data
The data is given as input
lines each containing 2
elements.
Initialize Counter
Word and Count
Extraction
After reading each line a
JON object, we extract
the title and the text
related to that piece of
news from it.
Create BoW
Dictionary
Create a dictionary and
add each word as the key
and its associated count
value as the value.

10. Counter Initialization
import sys
from collections import Counter
import json
bag_of_words = Counter()
Execution
for line in sys.stdin:
line = line.strip()
try:
word, count = line.split("t")
except:
continue
count = int(count)
bag_of_words[word] += count
with open('bow_data.json', 'w') as f:
json.dump(bag_of_words, f)

11. Moving to MongoDb
import json
from pymongo import MongoClient
client = MongoClient('mongodb://localhost:27017')
db = client['bow']
collection = db['bow_collection']
with open('bow_data.json', 'r') as f:
bow_data = json.load(f)
collection.insert_one(bow_data)
Performing MapReduce Operation
In the Terminal:
cat db.json | python3 bow_mapper.py | sort | python3 bow_reducer.py

12. HDFS
In the case of dealing with big data, we
could partition our dataset into a number
of batches instead of saving it in a single
file.
Instead of:
multiline_dataframe.write.save('/usr/local/hadoop/user3/dsba1.j
son', format='json')
Use:
partitioned_df = multiline_dataframe.repartition(4, "Unnamed: 0")
partitioned_df.write.save('/usr/local/hadoop/user3/dsba1.json',
format='json')
partition_counts = partitioned_df.rdd.mapPartitions(lambda it:
[sum(1 for _ in it)]).collect()
print(partition_counts)
[482, 480, 519, 519]

13. Create
Database
Create a database for
containing the data.
Import From
Hadoop
Import the JSON File
from Hadoop via
PySpark.
View Data &
Backup
View the data and if it is
inserted correctly then
create a backup before
starting the modifications.
Clean Data
Remove
non-alphanumeric
characters.
Display
Modified Data
Display the modified
content to view
changes.
MongoDB

14. Creating Database
Use an existing database or create a new one:
>use dsdb_dev
Viewing Data
>use dsdb_dev
>show collections
>db.fake_real_news.find()
>db.fake_real_news.aggregate([{$group : {_id: "$label", rest_number
: {$sum : 1}}}])
Creating a Copy
In the Terminal:
mongodump --db dsdb_dev --collection fake_real_news --out
/home/ds/Documents/
Importing From Hadoop
In the Terminal:
mongoimport --db dsdb_dev --collection fake_real_news --file
/usr/local/hadoop/user3/dsba1.json/part-00000-d1623440-4fde-4b
72-b87d-5943bec596d3-c000.json

15. Importing from Hadoop Using PySpark
with open('sampled_data.json', 'w') as file: for line in json_data: file.write(line +
'n')
import json
with open('sampled_data.json') as file:
data = file.readlines()
collection.insert_many([json.loads(line) for line in data])
df =
spark.read.json("/usr/local/hadoop/user3/dsba1.json/part-00000-d16234
40-4fde-4b72-b87d-5943bec596d3-c000.json")
sampled_df = df.sample(fraction=0.8, seed=42)
from pymongo import MongoClient
conn = MongoClient()
db = conn.dsdb_dev
collection = db['sampled_data']
json_data = sampled_df.toJSON().collect()

16. Data Cleaning
>db.fake_real_news.aggregate([
{‘$project': {‘_id': 1, 'Unnamed: 0': 1, 'label': 1, 'text': 1, 'title': 1} }
]). forEach(function(doc) {
if (doc.title) {
var newTitle = doc.title.replace(/[^a-zA-Z0-9 ]/g, '');
db.fake_real_news.update({ '_id': doc._id }, { '$set': { 'title': newTitle } });
}
});
Modified Content Display
>db.fake_real_news.aggregate([
{‘$project': {‘_id': 1, 'Unnamed: 0': 1, 'label': 1, 'text': 1, 'title': 1} }
]);
The file is now ready for word occurrence counting,
which can be done using Jupyter Notebook and
PyMongo.
Backup Restoration
In case of any need, restore the initial file:
>db.fake_real_news.drop()
mongorestore --db dsdb_dev --collection fake_real_news
/home/ds/Documents/dsdb_dev/fake_real_news.bson

17. Count the Number of Words
db.fake_real_news.aggregate ([
{
'$match': {
'label': "0" # the condition for the 'label'
field to be 1
}
},
{
'$project': {
'words': {'$split': [{'$toLower': '$title'}, ' ']} #
Split the lowercase version of the title field into
an array of words
}
},
{
'$unwind': '$words' # Separate documents
for each word
},
{
'$group': {
'_id': {
'word': '$words', # Group by word field
and count
},
'count': {'$sum': 1}
}
},
{
'$project': {
# Project to return only word field, count,
and id
'word': '$_id.word',
'count': 1
}
},
{
'$match': {
'word': {'$ne': None}, # Exclude null or
non-existent values
}
},
{
'$match': {
'$expr': {'$ne': ['$word', '']} # Exclude
empty strings
}
},
{
'$sort': {'count':-1}
}
])

18. Hypotheses
H1
Generation of fake news shall be with
the help of stop words.
Metrics - Average number of stop
words in title shall be higher in fake
news.
H2
Real news shall be short and crisp in
order to generate easy value.
Metrics - Length of the fake news shall
be more than the real ones.

19. H1
We used NLTK to extract stop words from
the title column and compared the
averages between fake and real titles.
The hypothesis is false, as shown by the
figure: fake news (0) is less frequent than
real news (1).

20. H2
The hypothesis is true, as shown by the figures:
fake news (0) tends to be longer than real news
(1).

21. Insights on Data &
Pre-processing
To gain quick insights from the data, we
used word clouds for the titles overall and
for fake/real data.

22. Real News

23. Fake News

24. Null Values
The title column contains some null
values, which may cause issues in data
analysis or processing.
We need to fill the null values in the title
column to ensure accurate data analysis.

25. Text Normalization
To further prepare the data, we applied text normalization techniques, including converting
the title and text to lowercase and removing punctuation marks.

26. Classification Model
For the binary classification of the News, we have choose
Random Forest Classifier
Splitting of data in x and y variable and Test and train split of
the data has been performed with 77 & 33 size.
The bag of words has been performed to the text of the news
(X_train & X_test) and by removing the stop words in English.
The Label Y_train & Y_test has the class of the news (Fake = 0
& Real = 1 )
Now the train data is feed to the RandomForestClassifier with
500 trees and the model has been tested with the test data
and the model classification confusion matrix is below.

27. Thank You For Your Attention!