Network-aware recommendations in online social networks

摘要

Along with the rapid increase of using social networks sites such as Twitter, a massive number of tweets published every day which generally affect the users decision to forward what they receive of information, and result in making them feel overwhelmed with this information. Then, it is important for this services to help the users not lose their focus from what is close to their interests, and to find potentially interesting tweets. The problem that can occur in this case is called information overload, where an individual will encounter too much information in a short time period. For instance, in Twitter, the user can see a large number of tweets posted by her followees. To sort out this issue, recommender systems are used to give contents that match the user's needs.udududThis thesis presents a tweet-recommendation approach aiming at proposing novel tweets to users and achieving improvement over baseline. For this reason, we propose to exploit network, content, and retweet analyses for making recommendations of tweets.ududThe main objective of this research is to recommend tweets thatudare unseen by the user (i.e., they do not appear in the user timeline) because nobody in her social circles published or retweetedudthem. To achieve this goal, we create the user's ego-network up to depth two and apply theudtransitivity property of the emph{friends-of-friends} relationship to determine interesting recommendations. After this step, we apply cosine similarity and Jaccard distance as similarity measures for the candidate tweets obtained from the network analysis using bigrams. We also count the mutual retweets between the ego user and candidate users as a measure of shared similar tastes. The values of these features are compared together for each of the candidate tweets using pairwise comparisons in order to determine interesting recommendations that are ranked to best match the user's interests.ududExperimental results demonstrate through a real user study that our approachudimproves the state-of-the-art technique. In addition to the efficiency of our approach in finding relevant contents, it is also characterized by the fact of providing novel tweets, which solves the over-specialization challenge or serendipity problem that appears when using content-based recommender systems as a stand alone approach of recommendation.