DETECTION AND DYNAMICS OF CYBERBULLYING IN ONLINE SOCIAL NETWORKS

ABSTRACT

Cyberbullying, defined as bullying perpetrated through the use of information technology, is a serious problem among children, adolescents and young adults. In this thesis, we study the detection of cyberbullies and the dynamics of cyberbullying in online social networks. In order to detect cyberbullies, we use learning models based on content features, user features and social network features. Understanding cyberbullying dynamics means predicting the effect of bullying comments on users with history of nonbullying. In the course of this study, we explore the spread of cyberbullying influence through the pairwise interactions between users. Using a dataset from MySpace, we first build a model for identifying bullies and non-bullies. We build a second model formulating the relationship between the influencer and the influenced user, predicting the dynamics by which a user with history of non-bullying turns into a bully.

Based on our experimental results, we find that content features are significant in detecting cyberbullies as well as cyberbullying dynamics. We find user features to be significant in detecting cyberbullies but not significant in detecting cyberbullying dynamics. We find social network features not to be significant in detecting cyberbullies but to be significant in detecting cyberbullying dynamics.

Our models can provide moderators and network administrators an effective way to identify cyberbullies as well as develop informed insights into the dynamics of cyberbullying.

 

Chapter 1 Introduction

Cyberbullying, defined as using information technology to willfully and repeatedly hurt, harass or threaten others [1] , is a serious problem among children, adolescents and young adults. According to recent statistics [2] , 52% of young people report being victims of cyberbullying. Other recent statistics [3] indicate that among more than 10,000 young people surveyed, 28% experienced cyberbullying on Twitter and 54% on Facebook. Compared with traditional bullying, cyberbullying is especially problematic because cyberbullying is not restricted by time and space [4] and can occur more frequently and intensely [5] , making it more difficult to identify and control.

Although cyberbullying occurs in the cyber environment, it can cause dire consequences in the real world. Cyberbullying can lead to frustration, depression, anger and social phobia among its victims [6] . More seriously, it can lead to suicide [7] . In 2008, a 13year-old British boy named Sam Leeson hanged himself in his bedroom because he was bullied on the social networking site Bebo [8] . Furthering the problem, retaliatory behavior can cause victims to become bullies themselves [9] .

For the above reasons, it has become necessary to investigate approaches to detect and control cyberbullying behavior in social networks.

There have been a number of studies and industry tools developed to deal with this problem. Studies in the field of social sciences [10-13] analyze the different factors involved in cyberbullying as well as how to prevent cyberbullying through educating students and reporting issues immediately. Studies [14-25] in the field of computer and information sciences focus on using machine learning techniques to detect abusive and offensive content. Existing industry tools [26-28] allow students and parents to filter or report harmful contents or contacts. Also, these tools provide a convenient way for the involvement of parents and school officials, enabling them to better protect kids online.

In this thesis, we focus on methods for detecting cyberbullies and cyberbullying dynamics in online social networks. To detect cyberbullies, we use machine learning techniques on three types of features, namely content features, user features and social network features. Content features allow us to analyze bullies based on the characteristics of their posted comments. User features allow us to characterize bullies through their demographic information. And social network features allow us to investigate bullies through their location in the network. Detecting bullies can help moderators and network administrators directly catch bullies and take necessary action such as issuing warnings and closing accounts. We also use the same feature sets together with posting times (elapsed time between posts) to characterize cyberbullying dynamics. We regard cyberbullying dynamics as the spread of cyberbullying through the pairwise interactions between users. Detecting cyberbullying dynamics means predicting whether users will be influenced by another user’s cyberbullying comment. We look into the interaction between influencers and influenced users by which a user with history of non-bullying observes a peer engaging in bullying and follows suit. Predicting the dynamics of cyberbullying can help moderators and network administrators more effectively prevent the spread of cyberbullying behavior.

We use Decision Tree C4.5 with 10-fold cross-validation based on different feature combinations to detect cyberbullies and characterize the dynamics of cyberbullying. For detecting cyberbullies, as expected, we find that content features are the most significant and social network features are the least significant. For detecting cyberbullying dynamics, we find that both content features related to the influencer (i.e. how offensive he or she is) and social network features are significant indicators. Social network features appear to indicate the importance of the role of the bullying user in the community, and his or her visibility in the network. Apparently, our results suggest that the more the user is central and key to the network, the more he or she will influence others to follow his or her reprehensible behavior.

The structure of this paper is as follows. In Chapter 2, we discuss related works. In Chapter 3, we introduce some basic concepts related to this study. In Chapter 4, we discuss the dataset as well as the analysis of three types of features. In Chapter 5, we focus on the experiments and results for detecting cyberbullies and cyberbullying dynamics. Chapter 6 is the conclusion of the study with pointers to future work.

DETECTION AND DYNAMICS OF CYBERBULLYING IN ONLINE SOCIAL NETWORKS. GET MORE  INFORMATION TECHNOLOGY PROJECT TOPICS AND MATERIALS

Sharing is caring!