Personalised Recommendation of Who to Follow Based on Fellowship of Followers Marko Adamko, Pavol Navrat and Alena Kovarova Faculty of Informatics and Information Technologies Slovak University of Technology in Bratislava, Slovakia markoadamko@gmail.com, pavol.navrat@stuba.sk and alena.kovarova@stuba.sk Abstract: This paper presents an approach to personalised recommendation of who to follow on Twitter. It is based on a simple observation that following the same Twitter account could be considered a fact that contributes to a possible similarity of followers. We proposed a simple scheme to elicit gradually candidates to follow which are most likely sources of microblogs that are of some interest to the user. We implemented a prototype and conducted experiments. They show that the basic idea works. We compared our system with the known “Who To Follow” application. Our system shows at least comparable performance and often gives better recommendations. Keywords: Personalised Recommendation, Microblog, Twitter, Information Stream. 1 INTRODUCTION Information overload is one of the significant characteristics of the present era. It manifests itself in many forms, one important of them being an information stream. The online social networking service Twitter administers the creation and sharing of microblogs called tweets. Users write and post very short (up to 140 characters) messages. Twitter distributes it to other users who subscribed to receiving messages from them by declaring themselves to be their followers. A user can also forward a received tweet (retweeting), which gives them the power to spread information broadly. In such a way, users receive streams of tweets. To put this into context, there are (as of July 2015) 645 million registered Twitter users, out of which 289 million are active ones, generating on average of 58 million tweets per day (Statistic Brain Institute, 2015). Each user faces a question: Which tweets from this huge stream ordered on a timeline should be a part of the specific stream that is shown to them? A user can influence contents of own stream by subscribing to certain accounts as their follower. If they are too cautious, subscribing to only very few accounts, the user may miss many interesting microblogs. If they follow everyone who potentially writes something interesting, they may be flooded by so many tweets that it becomes impossible to read them all in a given time. And even this scheme does not guarantee that some interesting tweets are not missed simply because their source has not been known to the user. It is desirable to identify which tweets are good to recommend (automatically) to the user, where these tweets have to be very close to their interests. Sources or clues to user’s interests are their twitting history and their social relations. We propose an approach to tweets recommendation that is based on identification of other accounts that are likely sources of interesting tweets. We devised a simple scheme that is able to recommend who to follow. Experiments show our approach performs similarly or better than a known solution. The structure of the rest of the paper begins with a brief commentary of related works. This is followed by our explanation of our approach to recommendation of tweet sources. Then we give evaluation and results; and finally, conclusions and future work are presented. 2 RELATED WORKS The present work falls within a broader context from several points of view. Connection between personalisation and user modelling has been intensively studied by many scholars (e.g. Barla 2011) and particularly in social media (Yin 2015). The role of a group of users in personalised recommendation has been stressed e.g. by Kompan (2013, 2014). Microblogs themselves are sources of valuable information and thus a subject to analysis aiming to identify opinions (Machova 2013) and sentiments (Korenek 2014) or to perform exploratory search (Zilincik 2013). J. Chen et al. (2010) experimented with recommending content from information streams. In designing a recommender, they explored various options. They contemplated a three dimensional design space: along the first dimension, various options of how to select candidate accounts; along the second one, various options of how to use content information; and along the third one, how to use social information. K. Chen et al. (2012) proposed a collaborative ranking model for recommending interesting tweets. The model collects preference information from many