On the Aggression Diffusion Modeling and Minimization in Twitter MARINOS POIITIS, Aristotle University of Thessaloniki, Greece ATHENA VAKALI, Aristotle University of Thessaloniki, Greece NICOLAS KOURTELLIS, Telefonica Research, Spain Aggression in online social networks has been studied mostly from the perspective of machine learning which detects such behavior in a static context. However, the way aggression diffuses in the network has received little attention as it embeds modeling challenges. In fact, modeling how aggression propagates from one user to another, is an important research topic since it can enable effective aggression monitoring, especially in media platforms which up to now apply simplistic user blocking techniques. In this paper, we address aggression propagation modeling and minimization in Twitter, since it is a popular microblogging platform at which aggression had several onsets. We propose various methods building on two well-known diffusion models, Independent Cascade (퐼퐶) and Linear Threshold (!) ), to study the aggression evolution in the social network. We experimentally investigate how well each method can model aggression propagation using real Twitter data, while varying parameters, such as seed users selection, graph edge weighting, users’ activation timing, etc. It is found that the best performing strategies are the ones to select seed users with a degree-based approach, weigh user edges based on their social circles’ overlaps, and activate users according to their aggression levels. We further employ the best performing models to predict which ordinary real users could become aggressive (and vice versa) in the future, and achieve up to 퐴*퐶=0.89 in this prediction task. Finally, we investigate aggression minimization by launching competitive cascades to “inform” and “heal” aggressors. We show that 퐼퐶 and !) models can be used in aggression minimization, providing less intrusive alternatives to the blocking techniques currently employed by Twitter. CCS Concepts: • Information systems ! Social networks; World Wide Web; • Computing methodolo- gies ! Network science; Modeling and simulation. Additional Key Words and Phrases: social networks, information diffusion, aggression modeling, aggression minimization, cascades, immunization 1 INTRODUCTION Online social media offer unprecedented communication opportunities, but also come with unfor- tunate malicious behaviors. Cyberbullying, racism, hate speech and discrimination are some of the online aggressive behaviors manifesting in such platforms and often have devastating consequences arXiv:2005.10646v5 [cs.SI] 30 Aug 2021 for individual users and the society as a whole. Aggression can be explicit through inappropriate posting such as negative feelings and embarrassing photos, or implicit when it unconsciously hurts online users (e.g., through negative gossip spreading). Overall, online social media users are often left exposed and vulnerable to potential aggression threats. Inter-disciplinary studies have focused on cyber aggression from the perspective of social psychology and computational sciences. Social learning and bonding [1, 30], as well as the theory of planned behavior [19] provide the basis of its theoretical formulation. Furthermore, machine learning has been used to detect such behavior in online platforms (e.g. [5, 10, 13, 38]). Even with all this body of earlier work, online aggression has not been uniformly defined [9]. Hence, online aggression is formulated under varying approaches, depending on the severity of the aggressive behavior, the type of platform and social interactions it facilitates, the power of the aggressor over the victim etc. Authors’ addresses: Marinos Poiitis, [email protected], Aristotle University of Thessaloniki, Thessaloniki, Greece; Athena Vakali, [email protected], Aristotle University of Thessaloniki, Thessaloniki, Greece; Nicolas Kourtellis, nicolas. [email protected], Telefonica Research, Barcelona, Spain. , Vol. 1, No. 1, Article . Publication date: August 2021. 2 Poiitis, et al. Interestingly, aggressive behavior has been found to be tightly related to the social circle of the individual expressing it [16], while its origins are located in the aggressive peer influence and the effect of the initial aggressors on their network [29], suggesting a potential cascading spreading trend from the source of the behavior to its connections. Additionally, the above findings have been expanded by analyzing the underlying network structure and the pair-wise interactions, further enhancing the intuition that online aggression propagates akin to a diffusion process. Therefore, the aggression’s overall effect can be strengthened as it propagates through the network [15]. Surprisingly, aggression propagation in the cyber-space has gathered little attention, primarily due to the complexity and dynamic nature of the problem. Specifically, the overall diffusion process is influenced by factors explicit to the specifics of each social network, i.e, how users form connections with each other, as well as network-agnostic ones such as the effect of user anonymity on the aggression manifestation [42]. This hardship is evident from the lack of automated processes in the popular media platforms, such as Twitter, to mitigate aggression’s negative effects, which are restricted to blocking or reporting abusive users and removing inappropriate content [2]. The present work focuses on the online aggression propagation problem and studies the complexi- ties of modeling and minimizing aggression propagation in online social networks. It particularly focuses on Twitter, which apart from being one of the largest social media platforms, it openly provides user data that have been already analyzed by several studies within the context of ag- gression. It considers the well known diffusion models of Independent Cascading (IC) and Linear Threshold (LT) [18] as they provide the building blocks for studying the diffusion process at its two fundamental types of interactions: user-user (퐼퐶) and user-neighborhood (!) ). To this end, we formulate aggression-aware information diffusion through appropriate parameters and enable a thorough study of the aggression dynamics. Then, we show how 퐼퐶 and !) models can be used in aggression minimization, providing less intrusive alternatives to the techniques currently in use by Twitter. Fig. 1. The overall process of the current work. Initially, aggression modeling is studied by setting and experimenting on its various sub-components, including selection of seed nodes, weighting scheme of edges, transfer of aggression and activation of nodes. The best models can then be used to minimize the overall network’s aggression with further configurations by applying healing strategies. Finally, the best methods are compared to the currently used blocking mechanisms. The overall process is outlined in Figure 1. Initially, aggression diffusion modeling is examined, where its sub-components - seed selection strategy, weighting scheme, activation criterion/threshold strategy and aggression transferring - are defined and explored. The process leads to two models, an 퐼퐶 and an !) -based, with the best expressive performance. Next, aggression minimization , Vol. 1, No. 1, Article . Publication date: August 2021. On the Aggression Diffusion Modeling and Minimization in Twitter 3 step exploits these two models to define and examine the method of competitive cascades andits dynamics (i.e., healing strategy) as an alternative to the blocking methods, which are currently used by Twitter, to reduce the overall aggression level of the social network. To address the above considerations regarding online aggression diffusion and minimization, the main contributions of this work are: (1) C1: aggression diffusion theoretical foundation, by exploring 퐼퐶 and !) as the basic propagation models. Upon them, the theoretical notions of user-user interaction, initial user selection strategies and user propagation are introduced and adjusted accordingly. The same models are also utilized in aggression minimization methods, where various healing approaches are formulated (Section 3). (2) C2: aggression modeling and minimization experimentation, for both 퐼퐶 and !) , through similarity performance tests, statistical validation as well as aggression reduction specifically for the minimization process. Real data experimentation, extensive simulations, a modeling case study and comparison to the blocking minimization methods currently in use by Twitter and similar online social platforms support our findings (Section 4). (3) C3: Exploration on results and implications, by validating the experimental results for both modeling and minimization. Specifically, neighborhood similarity is shown to bethe most appropriate criterion according to which user relationships are formed, while central users are the best to initiate the diffusion process. Finally, the adjusted competitive cascades are shown to outperform current Twitter banning methods achieving a reduction of ∼50% for 퐼퐶 and ∼15% for !) (Section 4). (4) C4: Simulation framework release for reproducibility purposes and further experimenta- tion or extensions1. The rest of this paper is organized as follows. In Section 2 the related works are reviewed. Section 3 provides the theoretical foundation of both aggression modeling and minimization. The experimental evaluation is presented in Section 4 and Section 5 provides a discussion of this work’s results and further improvements.