Self-tuning in Graph-Based Reference Disambiguation

Nowadays many data mining/analysis applications use the graph analysis techniques for decision making. Many of these techniques are based on the importance of relationships among the interacting units. A number of models and measures that analyze the relationship importance (link structure) have been proposed (e.g., centrality, importance and page rank) and they are generally based on intuition, where the analyst intuitively decides a reasonable model that fits the underlying data. In this paper, we address the problem of learning such models directly from training data. Specifically, we study a way to calibrate a connection strength measure from training data in the context of reference disambiguation problem. Experimental evaluation demonstrates that the proposed model surpasses the best model used for reference disambiguation in the past, leading to better quality of reference disambiguation.