| dc.description.abstract | The Fediverse is experiencing a growth in popularity, becoming a viable alternative to Twitter and other “centralised” social networks. The recent acquisition of Twitter by Elon Musk, which has sparked controversy, has further intensified this phenomenon. The Fediverse encompasses an expanding array of decentralised social networks, such as Pleroma and Mastodon, which all share the same social federation protocol, known as ActivityPub. Each of these decentralised social networks is composed of independent servers (aka ”instances”) run by different administrators. Users can create accounts on any given server, allowing them to use the platform. Importantly, servers can then “federate” together in a peer-to-peer fashion, allowing users to interact with individuals on other servers. This creates a physically decentralized system that underpins a global set of interconnections. The general concept of the Fediverse is appealing as it promises increased openness and freedom of speech. This level of freedom, however, also creates a conducive environment for the growth of toxic communities. There is therefore need for content moderation. This important task lies with specialized users on the instances referred to as administrators. Such users are responsible for managing all aspects of the server and community. In centralized platforms, there are ongoing initiatives to automate the moderation process, not to mention the large pool of paid human resources. On the contrary, instance administrators within the Fediverse frequently run their instances as volunteers, and the majority of their moderation activities are carried out manually. For example, this may involve reviewing posts and deleting any that contravene certain terms of service. Because of this, “amateur” platforms have developed strategies to make the procedure easier to perform, often sacrificing the quality of decision making. This thesis therefore argues that it would be beneficial to propose ways to inject some automation into the process. The above motivates the research. The major goal of this thesis is to understand how content moderation is carried out in the Fediverse, identify what challenges are faced, and propose solutions to these challenges. To achieve this goal, we first conduct a large-scale measurement campaign to collect data. As a case study of the Fediverse, the measurements focus on one platform called Pleroma. Pleroma is a microblogging platform that has comparable functionality to Twitter. We then characterize Pleroma with the aim of understanding how administrators moderate currently. We then identify and quantify the challenges with the current mode. Lastly, we create and develop proposed efficient moderation frameworks that are specifically customised to the decentralised structure of online social networks. This thesis therefore offers a comprehensive analysis of the Fediverse, focusing on the historical development and underlying principles of instance federation and the construction of the Fediverse. A key theme throughout the thesis is identifying major challenges that do not exist in prior ”centralised” solutions. Due to the unique characteristics of the Fediverse, there exist a variety of challenges that are inherent in the task of collecting data from Fediverse platforms. To address this, in Chapter 3, we propose a data collection approach for both our preferred platform, Pleroma, and a comparable platform, Mastodon. We then proceed to characterise our dataset and offer an understanding of the various user timelines that users can utilise for communication. Our categorization includes a significant discussion of the function of instance administrators in the Fediverse. Using the data, the thesis then focuses on one particular function in the Fediverse: content moderation. This is the process by which server administrators ensure that users obey the terms of service, e.g., removing hate speech or spam. The thesis offers insight into the different types of federation policies available to administrators, how administrators have applied these federation policies, and to which instances these policies are applied against. The thesis explores two problems in this domain. The first challenge that arises is the problem of “collateral damage”, explored in Chapter 4. Collateral damage refers to users who are impacted by the application of policies against their instance due to actions of other users on their same instance. For example, this happens when an entire instance is blocked due to the activities of just one person. We find evidence that over 95% of users on moderated instances experience such collateral damage. To address this concern, we make recommendations on the implementation of federation policies. Further, we develop a model capable of recommending certain instances to administrators that may require more nuanced moderation rather than an outright reject. It is possible to make accurate predictions (f1=0.76) regarding instances that require such finer grained moderation. One explanation for this collateral damage is that administrators do not have sufficient time to make fine-grained decisions regarding moderation. Thus, being overburdened, especially as volunteers, is another major challenge that administrators may encounter. Chapter 5 explore this issue. We find that instances are often “understaffed”, with the majority of instances only having a single administrator, and recruiting no other moderators to assist, despite many having over 100K posts. We investigate the extent of this overhead and examine potential strategies to mitigate its impact. We observe a diversity of administrator strategies, with evidence that administrators on larger instances struggle to find sufficient resources. We show that it is possible to predict (f1=0.85) which instances will have policies applied against them and design WatchGen, a tool that flags particular instances for administrators to pay special attention to. Pleroma is one example of a Fediverse platform and dataset collection in the Fediverse can be challenging. As a result, we believe it will be of significant benefit to be able to apply solutions built on one platform to another. For example, models built on data from Pleroma being applied to data from the Mastodon platform. In Chapter 6, we explore this possibility. For this investigation we use WatchGen from Chapter 5 to test this model transfer. We achieve promising results with an f1-score of 0.59 when applying WatchGen built exclusively on Pleroma data to Mastodon data. We believe this can be improved with more training data. The thesis is concluded in Chapter 7, bringing together the findings to highlight a clear set of remaining challenges. | en_US |
