Predicting Topic Emergence in Disinformation-Prone Domains Using Hybrid Feature Sets: Corpus Geometry, Textual Content, and Social Dynamics

Most disinformation studies analyze narratives only after they have become established and widely disseminated, when topics are clearly identifiable and fact-checking becomes feasible. In contrast, emerging narratives in news streams appear initially through fragmented signals and do not yet form coherent topics. In embedding-based topic models, such early signals typically manifest as outliers. However, outliers are heterogeneous. They may correspond to pure noise, isolated marginal discourse, topic drift, the early stages of new topic formation, or emerging discourse clusters associated with controversies. At the publication date, these cases are indistinguishable using unsupervised methods alone, making early topic emergence prediction ill-posed without learning from past topic formation processes. To resolve this ambiguity, we rely on retrospective evidence of how topics formed in the past.