Learning Parameters in Entity Relationship Graphs from Ranking Preferences

Semi-structured entity-relation (ER) data graphs have diverse node and edge types representing entities (paper, person, company) and relations (wrote, works for). In addition, nodes contain text snippets. Extending from vector-space information retrieval, we wish to automatically learn ranking function for searching such typed graphs. User input is in the form of a partial preference order between pairs of nodes, associated with a query. We present a unified model for ranking in ER graphs, and propose an algorithm to learn the parameters of the model. Experiments with carefully-controlled synthetic data as well as real data (garnered using CiteSeer, DBLP and Google Scholar) show that our algorithm can satisfy training preferences and generalize to test preferences, and estimate meaningful model parameters that represent the relative importance of ER types.