Predicting and Characterising User Impact on Twitter

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 405–413, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Predicting and Characterising User Impact on Twitter Vasileios Lampos1, Nikolaos Aletras2, Daniel Preot¸iuc-Pietro2 and Trevor Cohn3 1 Department of Computer Science, University College London 2 Department of Computer Science, University of Sheffield 3 Computing and Information Systems, The University of Melbourne v.lampos@ucl.ac.uk, {n.aletras,d.preotiuc}@dcs.shef.ac.uk, trevor.cohn@gmail.com Abstract The open structure of online social net- works and their uncurated nature give rise to problems of user credibility and influ- ence. In this paper, we address the task of predicting the impact of Twitter users based only on features under their direct control, such as usage statistics and the text posted in their tweets. We approach the problem as regression and apply linear as well as non- linear learning methods to predict a user impact score, estimated by combining the numbers of the user’s followers, followees and listings. The experimental results point out that a strong prediction performance is achieved, especially for models based on the Gaussian Processes framework. Hence, we can interpret various modelling com- ponents, transforming them into indirect ‘suggestions’ for impact boosting. 1 Introduction Online social networks have become a wide spread medium for information dissemination and inter- action between millions of users (Huberman et al., 2009; Kwak et al., 2010), turning, at the same time, into a popular subject for interdisciplinary research, involving domains such as Computer Sci- ence (Sakaki et al., 2010), Health (Lampos and Cristianini, 2012) and Psychology (Boyd et al., 2010). Open access along with the property of struc- tured content retrieval for publicly posted data have brought the microblogging platform of Twitter into the spotlight. Vast quantities of human-generated text from a range of themes, including opinions, news and everyday activities, spread over a social network. Naturally, issues arise, like user credibility (Castillo et al., 2011) and content attractiveness (Suh et al., 2010), and quite often trustful or appealing informa- tion transmitters are identified by an impact assess- ment.1 Intuitively, it is expected that user impact cannot be defined by a single attribute, but depends on multiple user actions, such as posting frequency and quality, interaction strategies, and the text or topics of the written communications. In this paper, we start by predicting user impact as a statistical learning task (regression). For that purpose, we firstly define an impact score function for Twitter users driven by basic account proper- ties. Afterwards, from a set of accounts, we mea- sure several publicly available attributes, such as the quantity of posts or interaction figures. Textual attributes are also modelled either by word frequen- cies or, more generally, by clusters of related words which quantify a topic-oriented participation. The main hypothesis being tested is whether textual and non textual attributes encapsulate patterns that affect the impact of an account. To model this data, we present a method based on nonlinear regression using Gaussian Processes, a Bayesian non-parametric class of methods (Ras- mussen …

Content truncated to 500 words