Propagation Structure as a Signal for Misinformation Detection with Graph Neural Networks and Edgeless Baselines (Technical Report)

Abstract

Propagation-based detectors classify a story from the shape of the cascade that carries it, and they report strong benchmark figures. This article measures what those figures are worth. On UPFD, a bidirectional graph convolutional network reaches macro-F1 0.832 on PolitiFact and 0.920 on GossipCop. Handed the same node features with every edge deleted, a logistic regression reaches 0.940 on the second, and one integer, the number of accounts, reaches 0.723. A paired bootstrap finds it ahead on one of four configurations, +0.103, behind on another, -0.021, and indistinguishable on the two smallest. The configuration it loses is one where the benchmark hands it the news article, and masking that article reverses the outcome on GossipCop while costing accuracy on PolitiFact, so what the article embedding is worth changes sign between corpora, by more than the architectural differences this literature reports. Macro-F1 is already at its level from a fifth of a cascade observed. An influence ranking over reach, PageRank and k-core answers negatively on real data: against a 14.5% chance baseline set by unequal cascade sizes, the hundred most influential accounts sit at 13.4%. Carried onto 400 cascades collected from Bluesky the detector collapses, training from scratch beats fine-tuning at every data budget, and a one-parameter rule on cascade size outscores it, 0.902 against 0.768. What a propagation figure is worth depends on a baseline that reads no edge and on a feature supplied at the root, neither of which the compared studies report.