Researchers in Israel are using machine learning to help human fact-checkers identify the sources of fake news online.

The team, based at Ben-Gurion University of the Negev, found that fact-checkers were overwhelmed — especially at key moments, such as a US presidential election — by the sheer proliferation of fake news.

Their solution was to zoom out and look at the bigger picture, namely who’s seeing what content and how it spreads across social-media networks.

They examined “exposure networks” – the ways in which networks of users are exposed to misinformation, even if they don’t actively share or engage with it.

That gave them a much better idea of how fake news propagates through social media.

“We know little about how successful fact-checkers are in getting to the most important content to fact-check,” said Nir Grinberg. He jointly led the research project with Prof. Rami Puzis at Ben-Gurion’s Department of Software and Information Systems Engineering.

“That prompted us to develop a machine learning approach that can help fact-checkers direct their attention better and boost their productivity.”

By leveraging exposure networks, the machine-learning model can guide fact-checkers to focus their efforts on the most likely sources of misinformation, improve their efficiency in detecting fake news and help them adapt more quickly to new tactics used by fake news creators.

And the model can accomplish all this accurately at less than a quarter of the cost of standard human fact-checking.

“The problem today with the proliferation of fake news is that fact checkers are overwhelmed,” says Grinberg. “They cannot fact-check everything, but the breadth of their coverage amidst a sea of social-media content and user flags is unclear.”

The team demonstrates, in a paper entitled “Leveraging Exposure Networks for Detecting Fake News Sources,” how this technology can recognize telltale signs of fake news sources, such as unusual patterns in the way they spread, specific audience demographics and certain characteristics of the content.

Based on these patterns, the system can then flag potential fake news sources for further investigation.

Maintaining lists of fake news sites – which can appear and disappear very quickly — is a costly and labor-intensive endeavor. The machine learning model considers the audience’s appetite for falsehoods, which tend to grow more robust over time.

Grinberg says the system will need more real-world training, and should never replace humans, but rather “greatly expand the coverage of today’s fact checkers.”

The machine learning approach can significantly lower the burden on fact-checkers and produce reliable results over time, he says.

The paper was presented in August at the KDD Barcelona Conference on Knowledge, Discovery and Data Mining.

Originally posted at israel21c.org