Stochastic bilevel interdiction for fake news control in online social networks

Social media platforms attempt to mitigate and control fake news, using interventions such as flagging posts or adjusting newsfeed algorithms, to protect vulnerable individuals. In this paper, we consider performing intervention actions on specific source nodes or user-user edges in social networks, under uncertain effectiveness of different intervention strategies. We model misinformation from malicious users to vulnerable communities using stochastic network interdiction formulations. Specifically, we minimize the expected number of reachable vulnerable users via stochastic maximum flow, and develop an alternative formulation for handling large-scale social networks based on their topological structures. We derive theoretical results for path-based networks and develop an approximate algorithm for single-edge removal on paths. We test instances of asocial network with 23,505 nodes, based on the IMDb actors dataset, to demonstrate the scalability of the approach and its effectiveness. Via numerical studies, we find that characteristics of removed edges change when intervention effectiveness is stochastic. Our results suggest that intervention should target on (i) a smaller set of centrally located edges with nodes that represent communities where regulatory actions are more effective, and (ii) dispersed edges with nodes where intervention has a high chance of failure.