東華大學圖書館 |

A Study of Social Media Trolls Via Graph Representation Learning.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	A Study of Social Media Trolls Via Graph Representation Learning./
作者:	Camacho, Albert Orozco.
面頁冊數:	1 online resource (131 pages)
附註:	Source: Masters Abstracts International, Volume: 85-05.
Contained By:	Masters Abstracts International85-05.
標題:	Public opinion. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30718449click for full text (PQDT)
ISBN:	9798380707718

A Study of Social Media Trolls Via Graph Representation Learning.
Camacho, Albert Orozco.

A Study of Social Media Trolls Via Graph Representation Learning. - 1 online resource (131 pages)

Source: Masters Abstracts International, Volume: 85-05.

Thesis (M.Sc.)--McGill University (Canada), 2023.

Includes bibliographical references

In modern times, social media platforms provide accessible channels for interacting and sharing information about important real-time events. However, these platforms are also regularly targeted by coordinated information attacks. Twitter has publicly released datasets of confirmed fake accounts under their Information Operations program to allow researchers to study these attacks. These accounts are commonly known as Internet trolls.In this thesis, we study three of the troll datasets from the Twitter Information Operations program, whose origin has been traced to Russia, China, and the Internet Research Agency (IRA). We first augment each dataset with a carefully sampled control group of active users engaged with similar content but not suspended by Twitter. This provides us with a rich set of online posts to study how state-backed trolls behave, how the troll activity fluctuates over time, and how these fluctuations compare to active users. In particular, we use graph representation learningto encode users' activities in each timestamp into a link prediction task. These learned representations are then used to contrast troll and active users. We show that, on average, the model has a more challenging time predicting the activity of trolls in two of our three datasets. The model also struggles to classify trolls against active users in these two datasets. However, in the third dataset, trolls are more predictable and easily distinguishable from active users.We hypothesize that troll sophistication might be related to whether they target local or global events. Finally, we discuss how these representations could help us better understand the activities and how they engage with active users. To show this, we group link embeddings into clusters, and within each cluster, we contrast the content generated by both categories of users. Although an automatic classification might not be possible for sophisticated trolls, learning user graph representations is, at least, helpful to understand their patterns of activities better.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2023

Mode of access: World Wide Web

ISBN: 9798380707718Subjects--Topical Terms:

531264
Public opinion.
Index Terms--Genre/Form:

542853
Electronic books.

A Study of Social Media Trolls Via Graph Representation Learning.
LDR:05924nmm a2200397K 4500 001 2364635
005 20231130105858.5
006 m o d
007 cr mn ---uuuuu
008 241011s2023 xx obm 000 0 eng d
020 $a 9798380707718
035 $a (MiAaPQ)AAI30718449
035 $a (MiAaPQ)McGill_j3860d25r
035 $a AAI30718449
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Camacho, Albert Orozco. $3 3705452
245 1 2 $a A Study of Social Media Trolls Via Graph Representation Learning.
264 0 $c 2023
300 $a 1 online resource (131 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Masters Abstracts International, Volume: 85-05.
500 $a Advisor: Rabbany, Reihaneh.
502 $a Thesis (M.Sc.)--McGill University (Canada), 2023.
504 $a Includes bibliographical references
520 $a In modern times, social media platforms provide accessible channels for interacting and sharing information about important real-time events. However, these platforms are also regularly targeted by coordinated information attacks. Twitter has publicly released datasets of confirmed fake accounts under their Information Operations program to allow researchers to study these attacks. These accounts are commonly known as Internet trolls.In this thesis, we study three of the troll datasets from the Twitter Information Operations program, whose origin has been traced to Russia, China, and the Internet Research Agency (IRA). We first augment each dataset with a carefully sampled control group of active users engaged with similar content but not suspended by Twitter. This provides us with a rich set of online posts to study how state-backed trolls behave, how the troll activity fluctuates over time, and how these fluctuations compare to active users. In particular, we use graph representation learningto encode users' activities in each timestamp into a link prediction task. These learned representations are then used to contrast troll and active users. We show that, on average, the model has a more challenging time predicting the activity of trolls in two of our three datasets. The model also struggles to classify trolls against active users in these two datasets. However, in the third dataset, trolls are more predictable and easily distinguishable from active users.We hypothesize that troll sophistication might be related to whether they target local or global events. Finally, we discuss how these representations could help us better understand the activities and how they engage with active users. To show this, we group link embeddings into clusters, and within each cluster, we contrast the content generated by both categories of users. Although an automatic classification might not be possible for sophisticated trolls, learning user graph representations is, at least, helpful to understand their patterns of activities better.
520 $a En temps modernes, les plateformes de medias sociaux pourvoient des canaux accessibles pour l'interaction et reaction aux informations sur des evenements importants en temps reel. Cependant, ces plateformes sont egalement regulierement la cible d'attaques coordonnees d'informations. Pour permettre aux chercheurs d'etudier ces attaques, Twitter a publie des donnees de comptes factices confirmes dans le cadre de leur programme d'Operations d'Information. Ces comptes sont communement appeles trolls Internet.Dans cette these, nous etudions trois ensembles de donnees de trolls issus du programme d'Operations d'Information de Twitter, dont l'origine a ete retracee en Russie, en Chine et a l'Internet Research Agency (IRA). Nous augmentons, d'abord, chaque ensemble de donnees avec un groupe temoin soigneusement echantillonne d'utilisateurs actifs impliques dans des contenus similaires, mais non suspendus par Twitter. Cela nous fournit un riche groupe de publications en ligne pour etudier comment se comportent les trolls finances par l'Etat, comment l'activite des trolls fluctue dans le temps et comment ces fluctuations se comparent a celles des utilisateurs actifs. Particulierement, nous utilisons l'apprentissage de representations des graphespour encoder les activites des utilisateurs a chaque instant, en une tache de prediction des liens. Ces representations apprises sont ensuite utilisees pour contraster les trolls et les utilisateurs actifs.Nous montrons que, en moyenne, le modele a plus de difficulte a predire l'activite des trolls dans deux de nos trois ensembles de donnees. Dans ces deux ensembles, le modele a egalement du mal a classer les trolls par rapport aux utilisateurs actifs. Neanmois, dans le troisieme ensemble de donnees, les trolls sont plus previsibles et facilement distinguables des utilisateurs actifs.Nous emettons l'hypothese que la sophistication des trolls pourrait etre liee a leur cible d'evenements locaux ou globaux. Enfin, nous discutons de la facon dont ces representations pourraient nous aider a mieux comprendre les activites et comment elles interagissent avec les utilisateurs actifs. Pour montrer cela, nous regroupons des vecteurs de liens en clusters de discussion et, dans chaque cluster, nous contrastons le contenu genere par les deux categories d'utilisateurs. Nous concluons que, bien que la classification automatique ne soit pas possible pour les trolls sophistiques, l'apprentissage de representations des graphes des utilisateurs est au moins utile pour mieux comprendre leurs patronsd'activites.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2023
538 $a Mode of access: World Wide Web
650 4 $a Public opinion. $3 531264
650 4 $a Study abroad. $3 3557623
650 4 $a False information. $3 3564362
650 4 $a Presidential elections. $3 3560858
650 4 $a Graph representations. $3 3560730
650 4 $a Pandemics. $3 3464080
650 4 $a Cybernetics. $3 527459
650 4 $a Computer science. $3 523869
650 4 $a Epidemiology. $3 568544
650 4 $a Web studies. $3 2122754
650 4 $a Political science. $3 528916
650 4 $a Social psychology. $3 520219
655 7 $a Electronic books. $2 lcsh $3 542853
690 $a 0800
690 $a 0984
690 $a 0766
690 $a 0646
690 $a 0615
690 $a 0451
710 2 $a ProQuest Information and Learning Co. $3 783688
710 2 $a McGill University (Canada). $3 1018122
773 0 $t Masters Abstracts International $g 85-05.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30718449 $z click for full text (PQDT)