Introduction

This project is based off the first part of the Fake News Challenge(FNC-1) where the goal is "to explore how AI...might be able to combat the fake news problem." Identifying fake news is a complex task that can be broken down into a few steps, with a potential first being the comparison of topics or, more precisely, stance detection, across myriad news organizations. The goal of this project is to classify the relationship between a body of text with a headline as agree, disagree, discuss or unrelated.

As we are working with string data, there is a lot of pre-processing and feature engineering to do.

Data Preprocessing

• removing punctuation
• lower casing all text
• removing stop words
• tokenizing
• stemming
• creating n-grams

Feature Engineering

• basic n_gram count ratios
• TF-IDF vectorization
• SVD
• word embeddings with Word2Vec using the Google News Corpus pre-trained weights
• sentiment features to assign polarity, using nltk Sentiment Analyzer with VaderSentiment

Modeling

Completed models include:

• Adaboost
• Gradient Boosing
• LSTM Deep Learning

Dependancies

• Scipy (pandas, numpy, matplotlib)
• Scikit-Learn
• NLTK
• Gensim
• Keras

I completed all work on a Google Compute Engine (GCE) VM. You can get up and running quickly with a jupyter connected VM on GCE following my post on Medium
Executing the preprocessing and feature engineering in one session will need ~800GB though, all data will be pickled for later use. I would however advise using numpy's .npz instead as it has quicker read/write times.

Future Work

• Improve model performance
  • Different preprocessing
  • Model parameters
  • Explore Feature importance - which help most and why?
• Model visualizations
• Build Convolutional Neural Network
• Begin real world application
  • Scrape news sites, archive data
  • Build a working record of sources and reliability