.PEFMJOHUIF4QSFBEPG'BLF /FXTPO5XJUUFS 5BJDIJ .VSBZNBɼ4IPLP8BLBNJZBɼ&JKJ "SBNBLJɼ 3ZPUB ,PCBZBTIJ /BSB*OTUJUVUFPG4DJFODFBOE5FDIOPMPHZ /"*45 5IF6OJWFSTJUZPG5PLZP 7th IC2S2 (2021)

2 l Increasing the number of access to news on social media l About two-thirds of American adults get news on social media l News from various doubtful sources has been spread by many users l Occasionally, fake news is spread. l Impact of Fake news l The spread of fake news on social media impacts on various fields; democracy, social media ecosystem, journalism, economics and etc. Background from: Pew Research Center from: https://news.northeastern.edu/2016/11/23/3qs-what-a-shame-how-to-filter-out-fake-news/

3 Example of fake news: Health • Various fake news related to the COVID-19 • Vitamin C cures the COVID-19 • 5G mobile networks spread COVID-19 From: https://k-tai.watch.impress.co.jp/docs/news/1246278.html Background

4 Research Objective Modeling the cascade of fake news on Twitter • Understanding the cascade of fake news could be useful in the application of fake news; detection and mitigation. • Model the re-share probability of fake news leveraging by post time and user profile (number of followers) Background

5 Research Objective Modeling the cascade of fake news on Twitter • Understanding the cascade of fake news could be useful in the application of fake news; detection and mitigation. • Model the re-share probability of fake news leveraging by post time and user profile (number of followers) Breaking News!! Discovery a huge cat Fake News Timeline Post by users 2020/01/01 12:00 The cut is huge！ http://.... Modeling Post probability Background

6 Hypothesis: the cascade of fake news Fake news cascade on social media is comprised of two stage cascades • First Cascade has the characteristics as ordinary news story • Second Cascade has the characteristics of correction because users recognize the falsity of the news item around a correction time !" Proposed Model First cascade (the characteristics of ordinary news) Second cascade (the characteristics of correction)

7 Base Model: Time-Dependent Hawkes Process l Hawkes Process: one of point process models whose defining characteristic is that they “self-excite” l Calculate the probability of the next event from past events and elapsed time. From: A Tutorial on Hawkes Processes for Events in Social Media Event timing Probability of event Modeling Proposed Model

8 Proposed Model Timeline Post on social media Modeling The probability of post between t, t + ∆% = λ % ∆% = ((%) ∑ ,:./0. 1, 2(% − %, ) Infection rate How many people remember? Base Model: Time-Dependent Hawkes Process (TiDeH) ((%): infection rate at time t 1, : number of followers at 4.5 post %, : time at 4.5 post 2: the function of how much do you remember

9 Proposed Model Timeline Post on social media Modeling The probability of post between t, t + ∆% = λ % ∆% = ((%) ∑ ,:./0. 1, 2(% − %, ) Infection rate How many people remember? ( % = 4 1 − 6 sin 2; <= % + >? @A./C 4the intensity of information 6: the relative amplitude of the oscillation >? : phase D : the characteristic time of popularity decay ((%): infection rate at time t 1, : number of followers at E.F post %, : time at E.F post 2: the function of how much do you remember Circadian Rhythm the freshness Base Model: Time-Dependent Hawkes Process (TiDeH)

10 Time (hour) ade ) cade ion) s Modeling Time (hour) Posting Activity ! " = $% " ℎ% " + $( (")ℎ( (") "+ Split the cascade at the correction time "+ $% " ℎ% " $( (")ℎ( (") Proposed Model

11 Time (hour) ade ) cade ion) s Modeling Time (hour) Posting Activity ! " = $% " ℎ% " + $( (")ℎ( (") "+ Split the cascade at the correction time "+ $% " ℎ% " $( (")ℎ( (") Proposed Model ℎ% " = , -:/012-3(/, /5) 6- 7(" − "- ) ℎ( " = , -:/5 1 /0 1/ 6- 7(" − "- ) from "+ until "+ 9the intensity of information :: the relative amplitude of the oscillation ;< : phase = : the characteristic time of popularity decay $(") Our hypothesis The first cascade has the characteristics of ordinary news, while second cascade has the characteristics of correction

12 Dataset • Recent Fake News (RFN) • Collect 7 fake news, which reported by “Politifact” and “Snopes” from March to May, 2019. • Each news had over 300 posts and kept posting over 36 hours from the initial post. • Fake News in Tohoku earthquake (Tohoku) • Collect 19 fake news related to Tohoku earthquake, which reported by Japanese news media* from March 12th to March 24th 2011. • Each news had over 300 posts and kept posting over 36 hours from the initial post. &WBMVBUJPO * https://blogos.com/article/2530/

13 &YQFSJNFOU &WBMVBUJPO Task: Predict the number of posts about fake news in the future Results: Proposed method achieved higher accuracy than other methods in two dataset. (100% in RFN and 89% in Tohoku) The smaller value is better

14 Verification of the correction time !" • Check the text around the correctio time !" • Verify whether the proposed model can properly estimate the correction time !" from text of posts • Count words that mean hoax or correction (e.g. fake, not true…) • Our hypothesis The first cascade has the characteristics of ordinary news, while second cascade has the characteristics of correction %JTDVTTJPO Blue line indicate the correction time Black line indicates the number of fake words

15 Verification of the correction time !" : Word cloud %JTDVTTJPO Example: Fake news “Turkey Donates 10 Billion Yen to Japan” • Compare the word cloud around the correction time (!" = 37) • Before &', words related to the news “pro-Japanese” appear. • After &', some user point out the news is “false rumor.”

16 Conclusion • Model posts of fake news on Twitter as two cascades comprised of ordinary news and correction of fake news • The proposed model achieves higher accuracy in the prediction task of number of posts. • Around the correction time !" , some user begin to point out the news is “false rumor.”