Micro-blogging services constitute a popular means of real time communication and information sharing. Twitter is the most popular of these services with 300 million monthly active user accounts and 500 million tweets posted in a daily basis at the moment. Consequently, Twitter users suffer from an information deluge and a large number of recommendation methods have been proposed to re-rank the tweets in a user’s timeline according to her interests. We focus on techniques that build a textual model for every individual user to capture her tastes and then rank the tweets she receives according to their similarity with that model.
In the literature, there is no comprehensive evaluation of these user modeling strategies as yet. To cover this gap, in this thesis we systematically examine on a real Twitter dataset, 9 state-of-the-art methods for modeling a user’s preferences using exclusively textual information. Our goal is to identify the best performing user model with respect to several criteria: (i) the source of tweet information available for modeling (ii) the user type, as determined by the relation between the tweeting frequency of a user and the frequency of her received tweets, (iii) the characteristics of its functionality, as derived from a novel taxonomy, and (iv) its robustness with respect to its internal configurations, as deduced by assessing a wide range of plausible values for internal parameters. Our results can be used for fine-tuning and interpreting text user models in a recommendation scenario in microblogging services and could serve as a starting point for further enhancing the most effective user model with additional contextual information.