Deep Learning Approaches to Fake News Detection

• Explored a number of core deep learning algorithms for their performance in detecting fake news.
• Employed data augmentation techniques to increase the size and quality of the training data set.
• Also made use of GloVe word vectors to enable the algorithms to learn the data.

Code and Resources Used

Python Version: 3.6.9
Libraries: tensorflow, keras, spacy (en_core_web_sm), nltk, numpy, pandas, pickle, sklearn, matplotlib, seaborn, math, os
GloVe: https://nlp.stanford.edu/projects/glove/
LIAR Dataset: https://github.com/thiagorainmaker77/liar_dataset

Approach

The main goal was building a fake news detection model which generalizes well to different settings. The selected approach is illustrated by the following figure:

Conclusions

• Using data augmentation we were able to increase prediction accuracy by 5.1% compared to a similar architecture that did not make use of data augmentation.
• The Gated Recurrent Unit (GRU) performed best out of the selected algorithms.
• In order to achieve state-of-the-art accuracies on the LIAR dataset it is necessary to also perform a metadata analysis.
  • 'Credit history' is the metadata parameter that needs to be included to reach the state-of-the-art.