Mining Influencers in the German Twittersphere

Abstract

This conference contribution presents a new sampling approach for large follow networks of influential Twitter accounts, discusses first key results of its application to the German Twittersphere, and benchmarks their representativeness.
Twitter plays a major role in public communication as a preferred medium for politicians, journalists, and celebrities to directly communicate with their audiences. The structures and dynamics on the platform stand exemplarily as a source of empirical evidence for recent theories about a networked public sphere (Bruns, 2008; Bruns & Highfield, 2016). Not only is it possible to identify characteristic communication structures regarding specific topics, such as polarised crowds or community clusters (Himelboim, Smith, Rainie, Shneiderman, & Espina, 2017), but also to investigate the macro-structures of whole national Twitterspheres, for example, of Australia (Bruns, Moon, Münch, & Sadkowsky, 2017) and Norway (Bruns & Enli, 2018). However, for Germany, a similar overview of a national Twittersphere is missing, which hinders the examination of such phenomena within a German context. This is mainly due to the restrictive access policies of Twitter and its data vendors that make the gathering of a comprehensive dataset, as has been done for the Australian and Norwegian cases, a costly and/or time-consuming task.
Therefore, this ongoing project tests a new sampling approach to generate a large sample of the follow network of the most influential accounts in the German Twittersphere. Instead of collecting the follow connections of all German-speaking Twitter accounts, the data collection process used in this project is based on an adaption of the rank degree method (Salamanos, Voudigari, & Yannakoudakis, 2017b). This graph sparsification method has been shown to facilitate the identification of the most influential nodes in epidemic network models with samples of only small subgraphs of the full network, while preserving characteristic network measures (Salamanos, Voudigari, & Yannakoudakis, 2017a). The algorithm, in short, crawls the network by following paths through nodes with the highest degree. As it only needs locally available information on the network structure, we have adapted and implemented it to drive a data mining approach using Twitter’s standard Search API. So far, we have collected a network sample of over 800,000 nodes (Figure 1). The collection is still ongoing and aims for a sample of one million nodes. To test the representativeness of the sample, this project compares these new sample data against an existing dataset that has been collected by Bruns et al. (2017), comprising account details of all publicly accessible Twitter accounts in 2016.

This project provides a first authoritative representation of the German Twittersphere which enables further research, reveals community structures of influencers in the German Twittersphere, and permits inferences about changes in follow behaviours after recommendation algorithms were introduced on Twitter. Furthermore, the tested collection approach, whose code will be published under an open source license, will enable researchers to easily collect follow network samples of other countries’ public spheres, based on languages and time zones, as well as around arbitrary Twitter accounts of interest.