Epidemiological Modeling Of News And Rumors On Twitter

Epidemiological Modeling of News and Rumors on TwitterFang Jin , Edward Dougherty †, Parang Saraf , Yang Cao †, Naren Ramakrishnan Department of Computer ScienceGenetics, Bioinformatics, and Computational Biology DepartmentVirginia Tech, Blacksburg, VA 24061†{jfang8, parang, ycao, naren}@cs.vt.edu, †edougherty@vt.edu ABSTRACTCharacterizing information diffusion on social platforms likeTwitter enables us to understand the properties of underlying media and model communication patterns. As Twittergains in popularity, it has also become a venue to broadcastrumors and misinformation. We use epidemiological models to characterize information cascades in twitter resultingfrom both news and rumors. Specifically, we use the SEIZenhanced epidemic model that explicitly recognizes skepticsto characterize eight events across the world and spanninga range of event types. We demonstrate that our approachis accurate at capturing diffusion in these events. Our approach can be fruitfully combined with other strategies thatuse content modeling and graph theoretic features to detect(and possibly disrupt) rumors.Categories and Subject DescriptorsH.2.8 [Database Management]: Database Applications—Data Mining; I.2.6 [Artificial Intelligence]: Learning—Knowledge acquisition; Parameter learningGeneral TermsExperimentation, PerformanceKeywordsSIS, SEIZ, Epidemiological modeling, Rumor detection.1.INTRODUCTIONOnline social networks have become a staging ground formodern movements, with the Arab Spring being the mostprominent example. Nine out of ten Egyptians and Tunisiansresponded to a poll indicating that they used Facebook toorganize protests and spread awareness. As a precautionary measure, governments have taken to blocking social networking websites, showcasing the importance of understanding this phenomenon.Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, orrepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.The 7th SNA-KDD Workshop ’13 (SNA-KDD’13), August 11, 2013,Chicago, United States.Copyright 2013 ACM 978-1-4503-2330-7 . 5.00.Interestingly, the role of social networks is not limited tohelping organize the activities of disruptive elements. Manykey government and news agencies have also begun to embrace Twitter and other social platforms to disseminate information. After the tragic 2013 explosions at the BostonMarathon, the FBI resorted to online social networks tobroadcast crucial information about the suspects. The viraldiffusion of information provided them with vital information about the suspects. At the same time it is well knownthat online activity on sites such as Reddit led to mistakenidentification of some individuals and the spread of severalrumors.We were motivated to apply the latest in epidemiologicalmodeling to understand information diffusion on Twitter, inrelation to the spread of both news and rumors. Epidemiological models provide a classical approach to study howinformation diffuses. These models typically divide the total population into several compartments which reflect thestatus of an individual. For instance, common compartments denote susceptible (S), exposed (E), infected (I), andrecovered (R) individuals. Individuals transit from one compartment to another, with certain probabilities that have tobe estimated from data. The simplest model, SI, has twostates; susceptible (S) individuals get infected (I) by one oftheir neighbors and stay infected thereinafter. While conceptually easy to understand, it is also unrealistic for practical situations. The SIS model is popular in infectious diseasemodeling wherein individuals can transition back and forthbetween susceptible (S) and infected (I) states (e.g., thinkof allergies and the common cold); this model is often usedas the baseline model for more sophisticated approaches.The SIR model enables individuals to recover (R) but is notsuited for modeling news cascades on Twitter since there isno intuitive mapping to what ‘recovering’ means. The SEIZmodel (susceptible, exposed, infected, skeptic) proposed byBettencourt et al. [1] takes the interesting approach of introducing an exposed state (E). Individuals in such a state takesome time before they begin to believe (I) in a story (i.e.,get infected). While the authors of [1] used this approach tomodel the adoption of Feynman diagrams by communitiesof physicists, our work explores their use in modeling newsand rumors on Twitter.The key contributions of this paper are: Our work is the first to employ the SEIZ model tomodel real Twitter datasets. We employ non-linearleast squares optimization of the underlying systemsof ODEs over tweet data, and demonstrate how this

model is better at modeling rumor and news diffusionthan the traditional SIS model. We analyze eight representative stories (four true eventsand four rumors) across a range of topics (politics, terrorism, entertainment, and crime) and over several geographic regions (USA, Mexico, Venezuela, Cuba, Vatican). While not an exhaustive list, this demonstratesthe wide applicability of the proposed model. We demonstrate the capability of the SEIZ model toquantify compartment transition dynamics. We showcase how such information could facilitate the development of screening criteria for distinguishing rumorsfrom real news happenings on Twitter.2.RELATED WORKRumor modeling.As far as we know, Daley [5] first proposed the similaritybetween epidemics and rumors using mathematical analysis.Some researchers have studied rumor propagation modelingin different network topologies [13, 22]; however, they do notprovide any discussion of propagation differences betweennews and rumors. Shah et al. [17] detect rumor sources innetwork using maximum likelihood modeling. In [2], Budaket al. prove that minimizing the spread of the misinformation (i.e., rumors) in social networks is an NP-hard problemand also provide a greedy approximate solution. Castillo etal. [3] delve into twitter content modeling, such as sentimentanalysis and hashtags to identify rumors, while Qazvinianet al. [15] try to address this issue using broader linguisticmethods, to learn possible features of rumor and determinewhether a twitter user believes a rumor or not. More related work appears in [7, 19]. Our goal is to develop anunderstanding of these processes using diffusion models.Information Diffusion.Significant work has gone into research on informationdiffusion on social media, e.g., see [4, 9, 16, 21]. Recently,Matsubara etc. [10] conducted research on the rise and fallpatterns of information diffusion, and managed to capturethe power-law fall pattern and periodicities inherent in suchdata. Gomez-Rodriguez et al. [6] built a cascade transmission model to track cascading process taking place over anetwork; they traced overall blogs and news for a one-yearperiod and found that the top 1000 media sites and blogstend to have a core-periphery structure.Epidemiological models.Mathematical modeling of disease spread not only provides vital information about the propagation of the disease in a human network, but also offers insight into thestrategies that can be used to control them. The classification of the human population into different groups formsthe basic premise of using epidemiological models for modeling information diffusion. The two widely used such modelsare SIR (Susceptible, Infected, Recovered) and SIS (Susceptible, Infected, Susceptible) models. Newman et al. [14]showed that a large class of standard epidemiological models, viz. the SIR models, can be solved exactly on a widevariety of networks, and confirmed the correctness of solutions with numerical simulations of SIR epidemics on networks. Kimura et al. [8] proposed the application of the SISmodel to study information diffusion where the nodes canbe activated multiple times. Zhao et al. [23] proposed anSIHR (Spreaders, Ignorants, Hibernators, Removed) rumorspreading model, with forgetting and remembering mechanisms to simulate rumor spreading in inhomogeneous networks. Xiong et al. [20] proposed a diffusion model with fourdifferent states: susceptible, contacted, infected, and refractory (SCIR) to identify the threshold value of the spreadingrate approaches almost zero. Bettencourt et al. [1] proposedthe SEIZ (susceptible, exposed, infected, skeptic) model tocapture the adoption of Feynman diagrams by using the publication counts after World War II. They extract the generalfeatures for idea spreading and estimate the idea adoptionprocess. Their result showed that the SEIZ model can fit thelong term idea adoption process with reasonable error, butdoes not demonstrate whether this model can be applied onlarge scale datasets, or whether can be applied on Twitter,where the story unfolds in real-time.3.DATASETSWe focus on twitter datasets that have reliable coverage ofthe events being studied; the volume of tweets ranges fromas low as 791 to nearly three orders of magnitude greater.As described in Table 1, the news and rumors studied weredrawn from a variety of regions and across a diversity oftopics. Data collection was aimed at gathering tweets highlyrelated to the events under study. We employed customizedsets of keywords and hashtags pertaining to each incident.Finally, date range restrictions were used to define relevanttweets for each event. It is also pertinent to note that thetweets analyzed spanned a variety of languages: English,Spanish, Italian, and Portuguese.3.1News topicsBoston Marathon Bombings. Two pressure cooker bombsexploded near the finish line of 2013 Boston Marathon onApril 15, 14:49:12 local time, killing three people and injuring more than 264 others. The FBI released photographsand surveillance videos on online social networks which spreadlike wildfire and provided crucial leads for identifying thesuspects1 .Pope Resignation. Pope Benedict XVI announced his resignation on the morning of February 11, 2013. In nearly 6centuries, this was the first time a pope has stepped downfrom his office. This news received reactions from all acrossthe world2 .Amuay Refinery Explosion. Propane and butane gasleakage caused an explosion at the Amuay refinery in Venezuelaon August 25, 2012 1:11 am local time. The blast killed 48people, injured 151 others and damaged 1600 homes3 .Michelle Obama at the 2013 Oscars. In the 2013 Oscarawards ceremony, a big surprise was the appearance of USfirst lady Michelle Obama for presenting the ‘Best Picture’award4 at-the-oscars-announces-be st-picture-winner/