Misinformation or “fake news” on social media has recently become a prominent phenomenon that has attracted enormous attention due to its significant societal and economic impact. Social media companies like Facebook or Twitter were challenged to provide technological answers to the problem. The majority of existing technological approaches for coping with misinformation detection are based on content analysis by means of natural language processing algorithms that lack the knowledge of political or social context or “common sense” for correct news interpretation.
At Fabula, we took a radically different approach to algorithmic misinformation detection. Instead of considering the content of the news, we look at the manner in which the news spread on social networks. There is a growing empirical evidence that “fake” and real news spread differently, making it possible to find characteristic patterns telling them apart. Our team has developed a new class of ML algorithms called “Graph Deep Learning” capable of learning such patterns from examples. The underlying core algorithms are a generalization of convolutional neural networks to graphs that have been developed in our group over the past five years. Our experimental results showed that our model achieves high fake news detection accuracy in a short span of time. Among the key advantages of Fabula technology compared to traditional content-based methods is that it is agnostic to the content and thus the language of the news, and much harder to beat by adversarial techniques, as it relies on the collective behavior of the social platform users.
The project was based on two ERC grants (Consolidator grant LEMAN as PoC grant GoodNews) that have led to the establishment of the startup company Fabula AI in 2018 for developing the technology commercially. The company was acquired by Twitter in June 2019. Our technology is currently being employed by Twitter to address various challenges related to the platform health.
