General :

Our work proposes a graph-based semi-supervised fake news detection method.

First, this repo is a python implementation of the paper

Semi-Supervised Learning and Graph Neural Networks for Fake News Detection by Adrien Benamira, Benjamin Devillers,Etienne Lesot, Ayush K., Manal Saadi and Fragkiskos D. Malliaros

published at ASONAM '19, August 27-30, 2019, Vancouver, Canada copyright space2019Association for Computing Machinery. ACM ISBN 978-1-4503-6868-1/19/08 http://dx.doi.org/10.1145/3341161.3342958}

Install guide

Requires python >= 3.6.

Configuration

Copy the config/config.default.yaml and rename the copy config/config.yaml.

This file will contain the configuration of the project.

To use sparse matrices

Uses last (not even released) tensorly sparse features.

Install master version of sparse:

$ git clone https://github.com/pydata/sparse/ & cd sparse
$ pip install .

Then use this version of tensorly:

$ git clone https://github.com/jcrist/tensorly.git tensorly-sparse & cd tensorly-sparse
$ git checkout sparse-take-2

Then place the tensorly-sparse/tensorly folder our project structure.

To use Transformer

Install master version of sparse:

$ git clone https://github.com/huggingface/pytorch-openai-transformer-lm.git

Then place the pytorch-openai-transformer-lm/ folder in our project structure under teh name transformer

Methods

Method for article embedding and construct the graph:

There are multiple choices : Method of the co-occurence matrix / embedding with GloVe (mean or RNN) / Transformer / LDA-idf.

Co-occurence matrix

method_decomposition_embedding can be parafac, GloVe, LDA or Transformer.

embedding:
  # Parafac - LDA - GloVe - Transformer -
  method_decomposition_embedding: parafac
  method_embedding_glove: mean 
  rank_parafac_decomposition: 10
  size_word_co_occurrence_window: 5
  use_frequency: No  # If No, only a binary co-occurence matrix.
  vocab_size: -1

GloVe

The embedding with glove : download GloVe nlp.stanford.edu/data/glove.6B.zip

There is 2 method of embedding: mean or RNN

paths:
  GloVe_adress: ../glove6B/glove.6B.100d.txt

embedding:
  # Parafac - LDA - GloVe - Transformer -
  method_decomposition_embedding: GloVe
  method_embedding_glove: mean  # mean or RNN
  use_frequency: No
  vocab_size: -1

Transformer

Git clone the project transformer-pytorch-hugging face, rename the file transformer and download the pre-trained model of OpenAI. Set the config path :

paths:
  encoder_path: transformer/model/encoder_bpe_40000.json
  bpe_path: transformer/model/vocab_40000.bpe

embedding:
  # Parafac - LDA - GloVe - Transformer -
  method_decomposition_embedding: Transformer
  use_frequency: No
  vocab_size: -1

LDA-idf

embedding:
  # Parafac - LDA - GloVe - Transformer -
  method_decomposition_embedding: LDA
  use_frequency: No
  vocab_size: -1

Word Mover Distance

The idea is instead of using the euclidean distance, we can use the WMD

Install Python packages:

spacy
wmd

Learning Methods

PyGCN

The pygcn lib used is: tkipf pytorch implementation.

PyAGNN

The pyagnn lib used is based on dawnrange pytorch implementation.

Results

Our pipeline is described in our report

Here is our result on the dataset

Credits to

Kipf, Thomas N and Welling, Max, Semi-Supervised Classification with Graph Convolutional Networks. https://github.com/tkipf/pygcn.
TensorLy: Tensor Learning in Python, Jean Kossaifi and Yannis Panagakis and Anima Anandkumar and Maja Pantic
Vlad Niculae, Matt Kusner for the word mover's distance knn.
Attention-based Graph Neural Network for semi-supervised learning,

bdvllrs/misinformation-detection-tensor-embeddings