Detecção de Fake News via Sinais Implícitos de Crowds: Uma Abordagem para Mitigar o Cold-Start no Cálculo da Reputação

The evolution of Digital Media for News Distribution has changed how people share content. Anyone can freely share information, including fake news (i.e., false information shared intentionally). In this scenario, fake news detection approaches have been proposed, with notable attention given to the one that uses crowd signals. Such approaches explore the collective sense by combining opinions (i.e., signals) of a high number of users (i.e., crowd), considering the reputations of these users regarding their capacity to identify fake news. Although promising, the crowd signals approaches have a significant limitation when many users in the crowds have not interacted with prior news. This lack of information leads to a cold-start problem when calculating those users’ reputations. The present work raises the following hypothesis: the performance of the crowd signals-based detection models can be improved if they mitigate the cold-start problem by inferring the users’ reputation based on the behavior those users would present in the face of prior news. This hypothesis is grounded on the fact that people tend to share news that seems familiar to their beliefs. In a search to validate the raised hypothesis, SCS, a crowd signals-based fake news detection method, is proposed. SCS considers similarities among texts (e.g., title and content) of past news to infer the reputation of users with the cold-start problem. BERT, a known Large Language Model (LLM) was used to provide embeddings that represent the texts. Preliminary experimental results demonstrate the effectiveness of the proposed method when compared to the crowd signals-based SOTA in detecting fake news.