This study applies statistical techniques to analyze digital violence in social networks. More precisely, two topics that have generated controversy are defined to analyze the tone of the language used by participants in social networks. In addition, techniques are proposed to identify provocative activities and to classify users according to their anonymity, visibility, and likelihood of being a "bot". Finally, targeted networks are constructed from Twitter conversations related to these two topics to analyze the interactions of the participants.
Mainly, the study helps to corroborate the premise that mathematical analysis can be a valuable tool for the analysis of interactions between social network users. Although the scope of the study does not allow for generalization, within the scope of the cases studied it is observed: (i) the more diverse the discussion, the less toxic the interaction; (ii) comments to news items are more toxic than Tweets; (iii) there are more non-toxic users than toxic ones, but toxicity is not concentrated in a few users; (iv) the use of hashtags, being a visible user and having a profile picture are associated with less toxic messages; (v) more toxic messages are presented from men to women than the other way around; (vi) the audience is passive when trying to calm aggressive discussions.
For the two selected case studies (" Political Group" and "Gender Violence in Digital Spheres"), 69,717 tweets and 10,400 historical comments were collected from the portals of the main digital media in the country. Additionally, information was extracted (id, number of followers, followed, favorites, among others) from 35,919 Twitter users who participated in the conversations on the topics of interest.
For this exercise, 1500 tweets and 500 comments were marked for each case. The tagging was done manually between three categories: toxicity, provocation and calmness. Subsequently, different classification models such as logistic regression, Naive Bayes, Boosted Trees and Support Vector Machines with linear kernel were trained and the level of toxicity, provocation and calmness were predicted for each message in the database. Due to ambiguity in defining and marking the level of provocation, the models reach a maximum of 0.76 area under the ROC curve for the best model. The "provocation” feature was the most difficult to classify reaching a maximum area of 0.66.
To enrich the analysis, anonymous users, users visible in the conversation and automated accounts ("bots") were identified. Different characteristics of Twitter user profiles were taken to segment them into different groups using a k-means algorithm. For example, to identify anonymous users, variables indicating whether the account is verified, whether the name used has a defined gender, whether it has geo-referencing enabled and whether it has changed its original profile image were used. Once the groups are identified, each one is manually assigned an anonymity rating according to the variables of each group.
The analysis allows us to identify the dynamics of conversations quantitatively, in four dimensions defined by the MinTIC: Context, Sender, Receiver and Audience. Some results to highlight are the following:
Context: More toxicity is observed in news comments than in tweets, the more diverse the discussion, the less toxic the interaction and the behavior of the networks associated with each case study is different: it seems to be associated with a dynamic of public debate in the case of the political group, and of conversational harassment in the case of gender violence in digital environments.
Transmitter: There are more non-toxic users than toxic ones, and toxic comments tend to use more exclamation points and capital letters, and fewer hashtags. Also, visible users tend to be less toxic.
Receiver: We observe more toxic messages from men to women than from women to men, and from men to men than from women to women. We also see that the most targeted users are more central to the network.
Audience: Very few calming messages are evident. In addition, the most toxic messages on average exhibit fewer retweets, but the difference with non-toxic messages is small. We also see that the most toxic moments of conversations were followed by moments with little interaction between users.
In Projects, you can learn about real applications of our services, use cases and examples of them.
