Detecting and reducing the spread of fake news using multimodal data

Abstract

An overwhelming number of true and fake news stories are posted and shared in social networks and users diffuse the stories based on multiple factors. In this thesis, I address the problem of detecting and reducing the spread of fake news or misinformation by leveraging the signals coming from different sources. The key idea is to adopt computational approach for modeling the social behavior, such as flagging or sharing, of users in the social network.

First, I propose a method of detecting fake news using a particular diffusion property called homogeneity. Diffusion of news stories from one user to another depends not only on the stories' content and the genuineness but also on the alignment of the topical interests between the users. In this paper, we propose a novel Bayesian nonparametric model that incorporates such homogeneity of news stories as the key component that regulates the topical similarity between the posting and sharing users' topical interests. The model extends hierarchical Dirichlet process to model the topics of the news stories and incorporates Bayesian Gaussian process latent variable model to discover the homogeneity values. I train the model on a real-world social network dataset and find homogeneity values of news stories that strongly relate to their labels of genuineness and their contents. Through experimental demonstration I show that the supervised version of HBTP predicts the labels of news stories better than the state-of-the-art neural network and Bayesian models.

Second, I propose an algorithm that, using the signals from the crowd, suggests certain news articles to be fact checked at an optimized time so that the number of people receiving misinformation is minimized. Given the uncertain number of exposures from news articles, the high cost of fact checking, and the trade-off between flags and exposures, such crowd-powered fact-checking procedure requires careful reasoning and smart algorithms. In this thesis, I first introduce a flexible representation of the fact-checking procedure using the framework of marked temporal point processes. Then, I develop a scalable online algorithm, CURB, to select which stories to send for fact checking and when to do so to efficiently reduce the spread of misinformation with provable guarantees. In doing so, a novel stochastic optimal control problem for stochastic differential equations with jumps needs to be solved. Throughout the experiment, I show that the suggested algorithm is able to effectively reduce the spread of fake news and misinformation.

Finally, I present a method to combine the fact-checking algorithm (CURB) with the fake-news detection model (HBTP). By this combination, I propose an online fact-checking algorithm that returns the fact-checking time using the crowd's temporal signals as well as the news article's homogeneity index that is inferred from its users' alignment of topical interests. Therefore, this method, which we term CURB + h, leverages 1) users' temporal dynamics such as the time stamps of theirs exposures, shares, and flagging events, as well as 2) the textual information (i.e., the title and the content of the news articles) to come up with an optimal fact-checking time aimed to reduce the total number of users that gets exposed to misinformation because of the news article of interest. Using this method, the precision of correctly targeting the fake news increases by 65.4% and the total number of users that do not get exposed to misinformation in the presence of the fact-checking algorithm (i.e., misinformation reduction) increases by 69.2% when compared to the fact-checking algorithm CURB.