DAISY LabSharif University of Technology
Publications

News Labeling as Early as Possible: Real or Fake?

Maryam Ramezani, Mina Rafiei, Soroush Omranpour, Hamid R. Rabiee

IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, 27 August 2019

arXiv

Abstract

The time gap between a news item's release and the detection of its label is a significant step towards broadcasting real information and avoiding fake news, but there is a trade-off between minimising that gap and maximising accuracy. We focus on accurate early labelling of news and propose a model that considers earliness both in modelling and in prediction, using recurrent neural networks with a novel loss function and a new stopping rule. Given the context of the news, we first embed it with a class-specific text representation, then use the available public profile of users and the speed of news diffusion for early labelling. Experiments on real datasets demonstrate effectiveness in both earliness and accuracy compared to state-of-the-art baselines.

In short

Detecting fake news accurately is easier the longer you wait, and useless if you wait too long. This paper makes that trade-off explicit: earliness enters both the loss function and the stopping rule, so the model decides when it has seen enough of a cascade to commit to a label.

The signals are the text of the news, the public profiles of the users spreading it, and the speed of its diffusion.

Cite this paper

@inproceedings{ramezani2019news,
  title = {{News Labeling as Early as Possible: Real or Fake?}},
  author = {Maryam Ramezani and Mina Rafiei and Soroush Omranpour and Hamid R. Rabiee},
  booktitle = {IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining},
  year = {2019},
  eprint = {1906.03423},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1906.03423}
}

Related work