A proximity measure for link prediction in social user-item networks

Recommendation systems based on historical action logs between users and items are usually formulated as link prediction problems for user-item bipartite networks, and such problems have been studied extensively in the literature. With the advent of on-line social networks, social interactions can also be recorded and used for predicting user´s future actions. As such, the link prediction problem based on the union of a social network and a user-item bipartite network, called a social user-item network in this paper, has been a hot research topic recently. One of the key challenges for such a problem is to identify and compute an appropriate proximity (similarity) measure between two nodes in a social user-item network. To compute such a proximity measure, in this paper we propose using a random walk with two different jumping probabilities toward different neighboring nodes. Unlike the simple random walk, our method is able to assign different weights to different paths and thus can lead to a better proximity measure by optimizing the two jumping probabilities. To test our method, we conduct various experiments on the DBLP dataset [21]. With a 3-5 year training period, our method performs significantly better than random guess in terms of minimizing the root mean squared error.