“Fake news” has become one of the most loaded terms on the internet, used for everything from swaying elections to undermining public trust. So, what if we could use artificial intelligence (AI) and various computer detection models to identify fake news automatically by examining such factors as certain key words, punctuation marks, headlines, and the websites linked to articles? That’s the purpose of groundbreaking research being conducted by Professor Sandip Kundu of the Electrical and Computer Engineering Department.
Kundu is the head of an international team of collaborators, whose research is described in a paper titled “Analysis of Fake News Classification for Insight into the Roles of Different Data Types.” The paper has been accepted for publication by the 16th Institute of Electrical and Electronics Engineers International Conference on Semantic Computing (IEEE ICSC), to be held virtually from January 26 to 28, 2022.
As Kundu explains about the research, his approach is for the first time looking under the hood of the existing state-of-the-art computer models and methodologies for detecting fake news to develop a much more effective detection system than now exists.
“Unfortunately,” says Kundu, “most fake news state-of-the-art classifiers operate as a black-box with no explanation for their predictions. This work aims to use ‘explainability’ methodologies to enable transparency into a state-of-the-art fake news classifier.”
To produce this pioneering classifier, Kundu and his research team utilized an existing fake news detection model called SAFE, which has achieved the best score on both of the “FakeNewsNet” datasets. Then the team combined SAFE with an existing detection technique called Layer-Wise Relevance Propagation, which takes advantage of feedforward neural networks architecture and has been previously applied successfully on other Natural Language Program problems.
Finally, Kundu and his colleagues also applied a more generic and computation-intensive methodology called “representation erasure” to better understand the importance of assorted input features.
“To the best of our knowledge,” says Kundu, “this is the first work to merge these techniques to increase model transparency.”
According to Kundu, “In this work, our goal is to apply post-hoc analysis to a fake news detection model to study the contribution of various data modalities and features towards classification. Specifically, we are trying to provide an answer for questions such as: How does a state-of-the-art model learn the fake news problem? Which features and modalities does it consider the most important?”
As one result, says Kundu, “We found that certain words, punctuation marks, headlines, and the websites linked to each article are best predictors of fake news. Our new analysis technique scores each one of these categories.”
Kundu notes that the main student contributor on this project is doctoral candidate Victor Ferreira from the Federal University of Rio de Janeiro in Brazil.
“Since my sabbatical at the Federal University of Rio de Janeiro (UFRJ) in 2015,” explains Kundu, “I have been supervising PhD students at UFRJ who typically spend a semester to a year at UMass. Felipe Franca, a professor emeritus at UFRJ, is his local advisor.”
Both Ferreira and Franca joined Kundu as contributing authors on the IEEE ICSC paper. (January 2022)