Predicting Friendship Links in Social Networks Using a Topic Modeling Approach

Abstract

In the recent years, the number of social network users has increased dramatically. The resulting amount of data associated with users of social networks has created great opportunities for data mining problems. One data mining problem of interest for social networks is the friendship link prediction problem. Intuitively, a friendship link between two users can be predicted based on their common friends and interests. However, using user interests directly can be challenging, given the large number of possible interests. In the past, approaches that make use of an explicit user interest ontology have been proposed to tackle this problem, but the construction of the ontology proved to be computationally expensive and the resulting ontology was not very useful. As an alternative, we propose a topic modeling approach to the problem of predicting new friendships based on interests and existing friendships. Specifically, we use Latent Dirichlet Allocation (LDA) to model user interests and, thus, we create an implicit interest ontology. We construct features for the link prediction problem based on the resulting topic distributions. Experimental results on several LiveJournal data sets of varying sizes show the usefulness of the LDA features for predicting friendships.