Phony News Observations (PhoNOs)

The Kaggle Competition

This project is the result of a Kaggle dataset that was created by Clement Bisaillon. The point is to develop a model which can differentiate between real and fake news articles.

What is in this repo?

Files	Description
true.csv	dataset containing real news
fake.csv	dataset containing fake news
real_or_fake.ipynb	Jupyter notebook containing data cleaning steps and models

Data Dictionary

The data consisted of two data sets -- fake and real .csv files each with 23481 rows and 21417 rows of data respectively. Each of these files contained news titles and body texts.

Columns	Description
title	The title of the article in question
text	The body of the article
subject	Category of news
date	Date that the article was written
category	True = 1 ; Fake = 0
weekday	The day of the week

Overall Thoughts

During our EDA this graph of the most common words for both fake an real news was generated.

A total of eight models were generated, five based on the title, and three based on the text. The three models types were Logistic Regression, Random Forest Classifier, and Extra Trees Classifier.

We immediately reached for logisitic regression, and it proved to be the right choice. Whether we used title or text as the modeling feature, logistic regression beat out the other models handily.

Conclusions

Methods	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6	Model 7	Model 8
PorterStemmer()	X			X	X
CountVectorizer()	X	X
TfidfVectorizer()			X	X	X	X	X	X
LogisticRegression()	X	X	X	X		X
RandomForestClassifier()					X		X
ExtraTreesClassifier()								X
Title as Feature	X	X	X	X	X
Text as Feature						X	X	X
Train Score:	0.9849	0.9840	0.9505	0.9563	0.9504	0.9886	0.9895	0.9819
Test Score:	0.9660	0.9529	0.9588	0.9643	0.9623	0.9910	0.9893	0.9824
Hyperparameters used in best score:
Estimator/Transformer	Hyperparameter	Set to:
-	-	-
TfidVectorizer()	stop_words	english
LogisticRegression()	max_iter	1000
LogisticRegression()	solver	liblinear