Popularity prediction based on interactions of online contents

Abstract

The interactive behavior of Web users often makes some online contents more popular than others. Thus the popularity of online contents can help us understand public interest and attention behind user interactions. Modeling and predicting the popularity of online contents is an important research topic and can facilitate many practical applications in different domains. Previous work on popularity modeling and prediction usually treat each online content separately, and neglect the interaction information between online contents, represented as interaction relations. In this paper, we explore the interaction relations between online contents, specifically competition and cooperation relations, for popularity prediction. We first define the interaction relations between different online contents and propose a method for the calculation of interaction information. We then apply the non-negative matrix factorization (NMF) technique to get a low dimensional representation of interaction features for online contents, which are used by classifiers for popularity prediction. We finally evaluate the proposed approach using two datasets from SinaWeibo (i.e., original tweets and topic hashtags). The experimental results show that interaction features alone can yield relatively good performance, and by incorporating interaction features into traditional feature based methods, our method can further improve popularity prediction results.