Identifying Implicit Enterprise Users from the Imbalanced Social Data

Abstract

Identifying the implicit enterprise users in social media enables the improvement of data quality for many applications like user profiling and targeted advertisement, as they register as ordinary users but act like enterprise ones and hence become the noises in the data. The recognition of implicit enterprise users confronts two challenges: (1) it needs to be handled quickly with little cost due to the very nature of preprocessing, and (2) it is necessary to deal with the highly skewed distribution of implicit enterprise users and ordinary users, which is about 1:10 in a social media site Sina Weibo in China. To the best of our knowledge, this problem is so far unexplored.

In this paper, we present an efficient class-imbalance learning framework which involves several types of new features from the users’ profile. Specifically, a cost sensitive learning strategy is designed to overcome the problem arising from the skewed data, and a set of novel features are extracted from the profile rather than the main contents to greatly reduce the overhead of crawling and processing the microblogs. We conduct extensive experiments on a real data set consisting of 2200 users (2000 ordinary users and 200 implicit enterprise users, respectively) in Sina Weibo. The results demonstrate that our method significantly outperforms the baselines by a large margin.