Tue, Aug 24, 2021 @ 11:30 AM - 01:00 PM
Yan Liu, Emilio Ferrara, Barath Raghavan, Fred Morstatter, Kimon Drakopoulos
Aug 24, Tuesday from 11.30 am to 1 pm.
Diffusion Network Inference and Analysis for Disinformation Mitigation
The proliferation of false and misleading information on social media has greatly reduced trust in online systems. Disinformation is largely aimed at influencing public opinion and social outcomes, and has been associated with reduced intent towards pro-social behaviors, denial of science and truth, and increased prejudices. In this thesis proposal, we address challenges in disinformation mitigation leveraging the content propagation or diffusion dynamics of disinformation on social media, through diffusion network analysis and inference techniques.
We propose techniques for early detection of disinformation contents, with a conditional generative model of social media responses to the content, leveraging historical responses to enrich semantic understanding of why content is labeled as disinformation. Secondly, we investigate how disinformation spreads and propose an unsupervised, generative model for detection of ma- licious coordinated efforts in the spread of disinformation, by inferring latent influence between accounts and collective group anomalous behaviors from observed account activities. Besides detection, we address challenges in network interventions to limit disinformation propagation and prevent viral cascades, by inferring diffusion dynamics of disinformation and legitimate contents from observed, unlabeled cascades with a mixture model of diffusion.
In the proposed future work, we focus on characterizing engagement with disinformation and conspiracy groups on social media. In the U.S. Election, we will evaluate whether Twitter's restriction on the QAnon conspiracy group was effective in limiting its activities with a regression discontinuity design for estimating causal effects of Twitter's intervention. In addition, to address the critical challenges in disinformation labeling, we propose to study methods in uncertainty sampling for active label refinement of social media posts, weakly-labeled based on news source credibility, towards building large-scale disinformation datasets, minimizing expensive human fact-checking efforts to collect disinformation labels. The outcome of this thesis proposal is to improve mitigation techniques and characterization of disinformation for timely identification and containment and to inform the need for robust platform measures.
Audiences: Everyone Is Invited
Contact: Lizsl De Leon