Classifying COVID-19 Related Meta Ads Using Discourse Representation Through a Hypergraph

Ujun Jeong, Zeyad Alghamdi, Kaize Ding, Lu Cheng, Baoxin Li, Huan Liu

Abstract: Despite Meta’s efforts to promote health information in the COVID-19 pandemic, the growing number of ads is making online content control extremely challenging. To effectively categorize the ads, this work investigates the major discourses shared across Meta ads with various categories related to COVID-19. We propose an interpretable classification model that captures common discourses in the form of keywords and phrases in ads. Particularly, we propose to use hypergraph to connect ads and discourses to capture their high-order interactions. Experiments on a curated Meta Ads dataset show that our model can provide subject-specific discourses and improve classification performance significantly.