Detecting User Interests on Twitter via Seed Set Expansion

A central machine learning problem in user interest modeling is to learn what topics users in a social network (such as Twitter) are interested in. Indeed, this is a critical part of understanding users’ behavior. If performed with reasonable precision, it can be exploited in a wide range of applications. Much of the previous work on this problem involves computationally expensive text processing – making the approach language dependent and, thereby, not scalable to international markets. In this work, we propose a novel graph-based approach to interest modeling that is language independent. In particular, we start with a list of people who are influencers and the topics they are Known-For (also called seed set). First, we expand these Known-For topics to other influencers. Later, we learn InterestedIn topics for all users by propagating the known for topics through the social graph. Numerical results show that on the Twitter social network consisting of over 250M users, we are able to grow a seed set of 55K labeled accounts into 88% interest coverage. Additionally, survey results verify that the precision of the detected topics at this coverage is as high as 80%.