Liesbeth Allein, Marie-Francine Moens and Domenico Perrotta
We present work on multimodal fake news detection which focuses on ethically integrating user information in text-based classifiers. Based on assumptions from the social sciences, our novel learning algorithm avoids model decisions relying on user profiling. Instead of simply using explicit user information in the decision process, the algorithm inspires a classifier with the article’s social context during optimisation. This is done by constraining its parameters on correlations between the article’s text and the tweets/profile descriptions of the users who spread the article online. Our algorithm is shown to encourage text-based classifiers to better discriminate between unseen fake and true news.