Predicting missing links in social networks with hierarchical dirichlet processes

We address the problem of predicting missing links for a social network in Social Media by using user activity data. We propose a simple and natural probabilistic model with latent features (traits) for simultaneously generating links and activities in the set of nodes, and present an efficient method of learning the model from the observed links and activities. In order to estimate the total number of latent features and the probability distribution of them for each node from the observed data, we incorporate a hierarchical Dirichlet process (HDP) into the model. On the basis of the learned model, we present a method of predicting missing links in the social network. We experimentally show by using synthetic data that the proposed learning method can estimate the link creation probabilities in good accuracy when there is an enough amount of training data. Moreover, using real and synthetic data, we experimentally demonstrate the effectiveness of the proposed link prediction method.