Fake news are the new plague of the 21st century. With the advent of social networks and the easy and quick access to information, this disease has become more and more common. Through retweets, shares or likes, a piece of fabricated information can in a few moments gain real credibility thanks to the common ”validation”. Similar to the tragedy of the common, each user is selfish and does not take the time to verify the sources, preferring to believe in this information that is often incredible, revolutionary and built to make the buzz. However, for several years now, various methods have been used on social networks to address these problems by detecting and removing problematic messages from the platform as quickly as possible. Our objective is to analyse the state of the art of these methods, to implement a graph-based solution and to see if adding node-level features helps to increase the prediction score. We will take the example of Twitter and model a graph for each tweet posted. Finally, we will use the a version of the FakeNewsNet dataset [5] which is one of the reference dataset for this type of task.
- UPFD Framework
- FakeNewsNet Paper
- Inductive Representation Learning on Large Graphs
- Graph Attention Networks
- Fast Graph Representation Learning with PyTorch Geometric
- Fake News Detection using Semi-Supervised Graph Convolutional Network
- Efficient Estimation of Word Representations in Vector Space
- BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
- Negative Log Likelihood Ratio Loss for Deep Neural Network Classification