Identifying Influencers with Ensemble Classification Approach on Twitter

The popular tweets are often originated by ordinary people involving in an event producing interesting or intriguing contents which often be picked up by users so-called influencers who have a mass number of followers. Targeting the users who cost the viral burst in social media among the enormous users producing messages every second is a difficult job. If such influencers can be identified in real time when trends are happening, some incidents can be prevented, protected and even drawn benefits from. Identifying influencers can be done with data classification techniques. Our research proposes a technique for identifying influencers on Twitter social media platform using extracted features from six months of collected tweets from November 2016 to March 2017. The proposed method contains two types of feature which are static and dynamic features. The static features include the weekday, the time period of the first retweet, and its author's numbers of interaction and relation consist of retweets of the hashtag, followers, followings, friends, favorites, and statuses. The dynamic features are the changes in the author's numbers of interaction and relation since the first retweet to a target time. The features were extracted according to a period of time according to Bursting State and bursting duration of 30, 60, and 120 minutes. Four ensemble classification algorithms: Bagging, Random Forest, Extra Trees, and Voting were applied. The results demonstrate that Bagging algorithm yields the best results in general for a duration of 120 minutes, and even identifying influencers in the separated topic domain, it also yields better predictive results.