Retweeting beyond expectation: Inferring interestingness in Twitter

Online social networks such as Twitter have emerged as an important mechanism for individuals to share information and post user generated content. However, filtering interesting content from the large volume of messages received through Twitter places a significant cognitive burden on users. Motivated by this problem, we develop a new automated mechanism to detect personalised interestingness, and investigate this for Twitter. Instead of undertaking semantic content analysis and matching of tweets, our approach considers the human response to content, in terms of whether the content is sufficiently stimulating to get repeatedly chosen by users for forwarding (retweeting). This approach involves machine learning against features that are relevant to a particular user and their network, to obtain an expected level of retweeting for a user and a tweet. Tweets observed to be above this expected level are classified as interesting. We implement the approach in Twitter and evaluate it using comparative human tweet assessment in two forms: through aggregated assessment using Mechanical Turk, and through a web-based experiment for Twitter users. The results provide confidence that the approach is effective in identifying the more interesting tweets from a user's timeline. This has important implications for reduction of cognitive burden: the results show that timelines can be considerably shortened while maintaining a high degree of confidence that more interesting tweets will be retained. In conclusion we discuss how the technique could be applied to mitigate possible filter bubble effects.