Structured Prediction of Generalized Matching Graphs

A structured prediction approach is proposed for completing missing edges in a graph using par- tially observed connectivity between n nodes. Unlike previous approaches, edge predictions depend on the node attributes (features) as well as graph topology. To overcome unrealistic i.i.d. edge pre- diction assumptions, the structured prediction framework is extended to an output space of directed subgraphs that satisfy in-degree and out-degree constraints. An efficient cutting plane algorithm is provided which interleaves the estimation of an edge score function with exact inference of the maximum weight degree-constrained subgraph. Experiments with social networks, protein-protein interaction graphs and citation networks are shown.