Article Preview
TopIntroduction
Microblogging systems have become a major trend as well as an important communication vector. In less than seven years, Twitter1 has grown in a spectacular manner to reach more than 500 million users in August, 20132 from which 54% are active users3 (posting at least a tweet per month). Other similar systems like Sina Weibo4, Identi.ca5 or Plurk6, to quote the largest, also exhibit dramatic growth. In such systems, the length of a published piece of news (called by post in the following) is limited to 140 characters which corresponds on average to 14.7 terms (Foster et al., 2011), so clearly greater than 4-5 terms adverts as reported in (König, Church, & Markov, 2009) but smaller than RSS items (Hmedeh et al., 2011), blogs (Ma & Zhang, 2007) or Web pages which have a size of 450 to 500 terms excluding tags (Levering & Cutler, 2006). The rationale for using such small messages is that these services were designed to be accessed with traditional cell phones through Short-Text Messaging protocol7.
In microblogging, a user, represented by his account, follows other accounts to be notified whenever they publish some information. Conversely, s/he becomes a publisher for the accounts that follow her/him, which results in the existence of a large social graph. One of the main differences with other online social networks is that in microblogging, the graph linking the users can be seen as an interest graph, since users follow each other if they have interest in each other posts. Those networks work on an all-or-nothing fashion, i.e. if a user A follows a user B, A will receive all of B’s posts.
Microblogging is also characterized by the heterogeneous nature of users. In Twitter for instance, there exist some high update frequency accounts (newspapers, tech-blogs and journalists) and others that publish less than one post a week. Moreover, there exist very popular accounts (e.g. news channel accounts like @FoxNews8 or famous accounts like @BarackObama9 with more than 45 million followers), and others with 0 or 1 follower.
For various reasons (security, advertisement, control policy …) these systems rely on a centralized architecture10. Each post published is received by the central system that forwards it to all the followers of the publishing account11. The central system is also responsible of the search feature (Busch et al., 2012). Since the most active accounts are generally the ones with the highest number of followers, the system must face a tremendous amount of posts to forward. For instance Twitter, which claimed in 2011 more than 200 million tweets a day, had to deliver daily over 350 billion tweets12. This traffic overload (especially for high-followed accounts) represents from an architectural point of view a scalability bottleneck.
On the follower's point of view, the amount of posts received from the accounts s/he follows, between 30 and 200 depending on the system considered (Kwak, Lee, Park, & Moon, 2010) loses the reader in the middle of long feeds of posts. This results in poor data readability and potentially loss of valuable information.
In a UK-based questionnaire over 587 participant, Bontcheva, Gorrell, & Wessels (2013) report that 33.9% of users perceive that they receive too many posts and 70.4% have found the task of locating the interesting/relevant posts amongst the others difficult. Also, 66.3% of the interviewed users have felt at some time that they can’t keep up with the amount of received tweets. A direct consequence of this phenomenon is the high dynamicity of the graph: to avoid flooding, users who follow active accounts tend to unsubscribe because they can't manage the continuous flow of posts as studied in (Kwak, Chun, & Moon, 2011; Kivran-Swaine, Govindan, & Naaman, 2011).