Repairing Noisy Graphs

Graphs are a flexible way to represent data in a variety of applications, with nodes representing domain-specific entities (e.g., records in record linkage, products and types in an ontology) and edges capturing a variety of relationships between these entities (e.g., an equivalence relationship between records in record linkage, a type-subtype relationship between types in an ontology). Often, the edges in this graph are noisy, in that some edges are missing (i.e., real-world relationships that do not have corresponding edges in the graph) and some edges are spurious (i.e., edges in the graph that do not have corresponding real-world relationships). Directly analyzing noisy graphs can lead to undesirable outcomes, making it important to repair noisy graphs. In this talk, we describe an approach that takes advantage of properties of real-world relationships and their estimated probabilities to ask oracle queries (an abstraction of crowdsourcing) to efficiently repair the noisy graphs. We illustrate this approach for the case of graphs that are unions of cliques (which is the case for record linkage) and graphs that are trees (which is the case for ontologies), and present theoretical and empirical results for these cases. This is joint work with Donatella Firmani, Sainyam Galhotra and Barna Saha.