A Bayesian Multiagent Trust Model for Social Networks

Addressing Misinformation in Online Social Networks: Diverse Platforms and the Potential of Multiagent Trust Modeling

“…The first is a Bayesian approach to reasoning about whether messages in social networks should be recommended to users or rejected, based on expected utility of those messages to users. Anchored by a model of partially observable Markov decision processes (POMDPs) and intended for environments where ratings of peers are known, the framework progressively reasons about three primary factors: similarity of the user to the rater (based on commonly rated items, in the past), credibility of the rater and the actual ratings that are provided [2]. We refer to this as the Sardana model (displayed in Appendix C).…”

“…Determining whether a tweet contains misinformation is a classification problem, so the final output should be a binary (yes or no) or a probability that describes how likely a tweet contains misinformation (on a scale from 0 to 1). We propose an algorithm that contains the strengths of previous models, by conditionally using those models depending on the tweet given, and then using parameters estimated by those models as observations for Sardana's POMDP model [2]. Our algorithm first selects a subset of users not including the current user.…”

“…In examining Reddit, we decided that several factors should influence the algorithm that reasons about potential misinformation, beyond the core considerations of ratings, credibility and similarity that are in focus for the model of Sardana [2]. We begin by observing that to determine whether some message is misinformation or not, the direct way would be to verify whether the message is true or not.…”

“…We examined some of the unique qualities of this network in proposing a particular algorithm for detecting misinformation. Our solution returns once more the approach of Sardana [2] but now explicitly attempts to integrate a reasoning about the central concern of author credibility, as well as developing a stand-in for post credibility, based on the kinds of group reactions that may be revealed in a network like Facebook. A final concern that we integrate, which seems a prevalent factor in this environment, is the use of links within posts and then the credibility of these links.…”

“…The second aim is to reflect on whether existing AI multiagent trust modeling algorithms can be effectively applied in order to distinguish valued content from more suspect posts which may be filtered from a user's stream. Towards this end, we examine some specific models that we have developed ourselves (a Bayesian predictive algorithm of trustworthy content [2], a data-driven classifier of low reputation messages that draws from analysis of discussions that follow each post [3], a multi-faceted personalized data analysis to weight the factors predicting trust relations between peers [4], and a process for identifying rumour spread within social networks that draws from crowdsourced labels of content, using expert labelling (originally designed for healthcare applications) [5]). The conclusion that we reach is that a combination of these methods are best to integrate into an effective procedure for addressing misinformation, and that while there are unique challenges of specific networking environments, there are in fact common concerns that can be leveraged for a concrete first step of the analysis.…”