A Collective Approach to Graph Identification


There is a growing amount of observational data describing networks— examples include social networks, communication networks, and biological networks. As the amount of available data increases, so has our interest in analyzing these networks in order to uncover (1) general laws that govern their structure and evolution, and (2) patterns and predictive models to develop better policies and practices. However, a fundamental challenge in dealing with this newly available observational data describing networks is that the data is often of dubious quality—it is noisy and incomplete— and before any analysis method can be applied, the data must be cleaned, missing information inferred and mistakes corrected. Skipping this cleaning step can lead to flawed conclusions for things as simple as degree distribution and centrality measures; for more complex analytic queries, the results are even more likely to be inaccurate and misleading. In this paper, we introduce the notion of graph identification, which explicitly models the inference of a “cleaned” output graph from a noisy input graph. We show how graph identification can be thought of as a series of probabilistic graph transformations. This is done via a combination of component models, in which the component models construct the output graph by merging nodes in the input graph (entity resolution), adding and deleting edges (link prediction), and labeling nodes (collective classification). We then present a simple, general approach to constructing local classifiers for predicting when to make these graph modifications, and combining the inferences into an overall graph identification framework. The problem is extremely challenging because there are dependencies among the transformation; ignoring the dependencies leads to sub-optimal results and modeling the dependencies correctly is also non-trivial. Graph identification is closely related to work in information extraction [12]; information extraction, however, traditionally infers structured output from unstructured data (e.g., newspaper articles, emails), while graph identification is specifically focused on inferring structured data (i.e., the cleaned graph) from other structured data (i.e., the noisy graph, perhaps produced from an information extraction process). There is significant prior work exploring the components of graph identification individually; representatives include work on collective classification [7, 5, 6, 13], link prediction [4, 10, 8], and entity resolution [1, 2, 14]. More recently, there is work that looks at various ways these tasks are inter-dependent and can be modeled jointly [15, 11, 16, 9, 3]. To our knowledge, however, previous work has not formulated the complex structured prediction problem as interacting components which collectively infer the graph via a collection of probabilistic graph transformations. In addition to defining the problem and describing a component solution approach, we present a complete system for graph identification. We show how the performance of graph identification is sensitive to the intraand inter-dependencies among inferences. We evaluate on two real-world

Cite this paper

@inproceedings{Getoor2010ACA, title={A Collective Approach to Graph Identification}, author={Lise Getoor}, year={2010} }