Studying Abusive Behavior of #Gamergate on Twitter
Total Page:16
File Type:pdf, Size:1020Kb
Hate is not Binary: Studying Abusive Behavior of #GamerGate on Twier Despoina Chatzakouy, Nicolas Kourtellisz, Jeremy Blackburnz Emiliano De Cristofaro], Gianluca Stringhini], Athena Vakaliy yAristotle University of Thessaloniki zTelefonica Research ]University College London [email protected], [email protected], [email protected] [email protected], [email protected], [email protected] ABSTRACT understanding is crucial to enable eective mitigation and help Over the past few years, online bullying and aggression have be- social network operators to detect and block these accounts. come increasingly prominent, and manifested in many dierent Roadmap. In this paper, we start addressing this gap by perform- forms on social media. However, there is little work analyzing the ing a large-scale comparative study of abusive accounts on Twitter, characteristics of abusive users and what distinguishes them from aiming to understand their characteristics and how they dier from typical social media users. In this paper, we start addressing this typical accounts. We collect a large dataset of tweets related to the gap by analyzing tweets containing a great large amount of abusive- Gamergate (GG) controversy [3], which after two years since its ness. We focus on a Twitter dataset revolving around the Gamergate start has evolved into a fairly mature, pseudo-political movement controversy, which led to many incidents of cyberbullying and cy- that is thought to encompass semi-organized campaigns of hate beraggression on various gaming and social media platforms. We and harassment by its adherents, known as Gamergaters (GGers), study the properties of the users tweeting about Gamergate, the against women in particular. Then, we explore the dierences be- content they post, and the dierences in their behavior compared tween the GG-related accounts identied as abusive, and random to typical Twitter users. Twitter accounts, investigate how these dierences lead to dispro- We nd that while their tweets are often seemingly about aggres- portional suspension rates by Twitter, and discuss possible causes sive and hateful subjects, “Gamergaters” do not exhibit common of these dierences. We also look at accounts of users that were expressions of online anger, and in fact primarily dier from typ- deleted by their owner and not by Twitter. Further, we cluster GG ical users in that their tweets are less joyful. They are also more accounts that exhibit similar behavior, aiming to identify groups engaged than typical Twitter users, which is an indication as to of similar accounts that should have been suspended by Twitter how and why this controversy is still ongoing. Surprisingly, we but are instead still active. Based on the ndings of our clustering, nd that Gamergaters are less likely to be suspended by Twitter, we reason about what may have driven Twitter to not suspend thus we analyze their properties to identify dierences from typical them. Finally, we test the performance of a supervised method to users and what may have led to their suspension. We perform an automatically suspend Twitter users based on the various features unsupervised machine learning analysis to detect clusters of users analyzed. who, though currently active, could be considered for suspension since they exhibit similar behaviors with suspended users. Finally, Findings. In summary, we discover that users involved in Gamer- we conrm the usefulness of our analyzed features by emulating the gate were already-existing Twitter users probably drawn to the Twitter suspension mechanism with a supervised learning method, controversy, which might be the reason why GG exploded on Twit- achieving very good precision and recall. ter in the rst place. While the subject of their tweets is seemingly aggressive and hateful, GGers do not exhibit common expressions arXiv:1705.03345v1 [cs.SI] 9 May 2017 of online anger, and in fact primarily dier from random users in 1 INTRODUCTION that their tweets are less joyful. We nd that despite their clearly Abuse on social media is becoming a pressing issue. Over the past anti-social behavior, GGers tend to have more friends and followers few years, social networks have not only been targeted by bots than random users and being more engaging in the platform may and fraudsters [1, 30, 35], but have also been used as a platform for have allowed this controversy to continue until now. Surprisingly, harassing and trolling other individuals [28]. Detecting and mit- we nd that GGers are disproportionally not suspended from Twit- igating such activities presents important challenges since abuse ter in comparison to random users, which is rather unexpected performed by human-controlled accounts tends to be less homoge- given their hateful and aggressive postings. Suspended GG users neous than the one performed by bots, making it hard to identify expressed more aggressive and repulsive emotions, oensive lan- the characteristics that distinguish them from non-abusive attacks guage, and interestingly, more joy than suspended random users, (and detect them). Recent work showed that human-controlled and their high posting and engaging activity may have delayed accounts involved in harassment actually present degrees of syn- their suspension from Twitter. Also GGers who deleted their ac- chronized activity [15]. However, no systematic measurement has count demonstrated the most activity in comparison to other users been performed to understand what distinguishes a social network (deleted or suspended), exhibited signs of distress, fear, and sad- account behaving in an abusive way from a typical one. Such an ness. They have probably showed these emotions through their high posting activity lled with anger, reduced joy, and negative user/activity based attributes, e.g., number of friends/followers and sentiment. Such users have small social ego-networks which may users’ account age. Our work aims to use such attributes to study have been unsupportive or too small to help them before deleting and understand the dierent behavioral patterns between random their accounts. and Gamergate Twitter users, while shedding light on how such dierences aect their suspension and deletion rates on Twitter. Paper Organization. The rest of the paper is organized as follows. The next section reviews related work on measuring abusive be- Analysis of Gamergate. To create an abuse-related dataset, i.e., a haviors on online platforms. Section 3 introduces our dataset and dataset containing abusive behavior with high probability, previ- the steps taken for cleaning and preparing it for analysis, then, in ous works rely on a number of words (i.e., seed words) which are Section 4, we analyze the behavioral patterns exhibited by GGers, highly related with the manifestation of abusive/aggressive events. and compare them to random Twitter users. In Section 5, we discuss In this sense, a popular term that can serve as a seed word is the how users get suspended on Twitter, dierences observed between #GamerGate hashtag which is one of the most well documented GGers and random users, reasons for deviating from the expected large-scale instances of bullying/aggressive behavior we are aware rates, and a basic eort to emulate Twitter’s suspension mechanism. of [21]. The Gamergate controversy stemmed from alleged im- In Section 6 we discuss our ndings and conclude. proprieties in video game journalism, which quickly grew into a larger campaign centered around sexism and social justice. With individuals on both sides of the controversy using it, and extreme 2 RELATED WORK cases of bullying and aggressive behavior associated with it (e.g., We now review related work on studying/detecting oensive, abu- direct threats of rape and murder), #GamerGate can serve as a rela- sive, aggressive or bullying content on social media sources. Chen tively unambiguous hashtag associated with texts that are likely to et al. [6] aim to detect oensive content, as well as, potential oen- involve abusive/aggressive behavior from a fairly mature, hateful sive users based on YouTube comments. Both Yahoo Finance [9, 24] online community. In [22], the author shows that #GamerGate can and Yahoo Answers [18] have been used as a source of information be likened to hooliganism, i.e., a leisure-centered aggression were for detecting hate and/or abusive content. More specically, [18] fans are organized in groups to attack another group’s members. studied a Community-based Question-Answering (CQA) site and Also, [12] aims to detect toxicity on Twitter, considering #Gamer- nds that users tend to ag abusive content in an overwhelmingly Gate to collect a sucient number of harassment-related posts. correct way. In this paper, we also study a number of abusive users involved Cyberbullying has also attracted a lot of attention lately, for in this controversy via #GamerGate. However, we are the rst to instance [2], [16] and [17] focus on Twitter, Ask.fm, and Instagram, investigate the attributes characterizing these users with respect respectively, to detect existing bullying cases out of text sources. [2] to their Twitter account status (active, suspended, deleted), and considers a variety of features, i.e., user, text, and network-based, to perform an unsupervised and supervised analysis of suspicious to distinguish bullies and aggressors from typical Twitter users. In users for possible suspension. addition to text sources, [17] also tries to associate an image’s topic (e.g., drugs, celebrity, sports,