Modeling Implicit User Relations with Information Propagation Graph for User Influence Prediction

With the popularity of social networks, people participate in all kinds of discussions and express their opinions in various social network platforms. The conventional way of information propagation, such as word-of-mouth or traditional media, has developed rapidly in social networks where more people can be influenced irrespective of their geo distance. In order to quantify each person’s influence in social network, most related works are based on friend or follow relationships and post content features for analysis. Without these explicit relationships, it’s hard to estimate user influence. In this paper, we build an implicit user influence model based on people who share and respond to topic-relevant articles in the past. Firstly, given a topic, we collect the social interactions such as replies, likes or dislikes between topic-relevant posts and comments. Secondly, using these implicit relationships, we can obtain each user’s behaviors, and post sentiments to construct information propagation graphs. Finally, we combine the sentiment polarity of each post with the implicit relations among users to estimate user influence. From our experimental results in online discussion forum, the proposed method can achieve comparable performance in finding high influence users for the same topic as compared with centrality measures. By combining author and replier influences, the best performance can be achieved with an average NDCG of 0.9653. This shows the potential of our proposed method for user influence prediction. Further investigation is needed to verify the performance in larger scale.