Joint Neighborhood Subgraphs Link Prediction

A crucial computational task for relational and network data is the "link prediction problem" which allows for example to discover unknown interactions between proteins to explain the mechanism of a disease in biological networks, or to suggest novel products for a customer in a e-commerce recommendation system. Most link prediction approaches however do not effectively exploit the contextual information available in the neighborhood of each edge. Here we propose to cast the problem as a binary classification task over the union of the pair of subgraphs located at the endpoints of each edge. We model the classification task using a support vector machine endowed with an efficient graph kernel and achieve state-of-the-art results on several benchmark datasets.