10 — Fake News in Social Networks

The model proposed in this paper demonstrates a mathematical way of modeling the spread and reaction to “fake news” as it exists on social media (modeled here as directed graphs). Each “person” represents internal memory with a gated recurrent neural network, and they attempt to pick a $\theta \in { 0, 1 }$ where a 1 means that the person deems a claim true, and a 0 means anything else. (In short, this means that the final “decision” of an agent is binary.)

The paper then uses several different types of adversaries: One that is ignorant of the readers of the “news”; one that knows the network positions of the readers (“agents”); and one that knows network position and the agent’s private signal (which correlates to the opinions of a news reader).

Some findings were intuitive; star nodes in the graph (many followers, each follower follows comparatively few others) had disastrous impacts on the accuracy of the decisions made (a celebrity retweeting false information, perhaps). Meanwhile, other findings were less intuitive and more mathematically nuanced (I won’t repeat them here because I’m not confident yet I fully understand them).

Central to the paper is the idea of representing an agent in a way that retains the complexity of internal decision-making without overgeneralizing. Using a RNN to make agent-level decisions of validity, claim the authors, closely models a Bayesian actor; but I imagine this gives the model greater variance in individual actors’ behaviors: I have no idea if that more closely or less closely models actual humans on a social network.