Quantifying and Ranking Bias in Social Networks

In recent years, social network analysis has gained popularity as a method for analyzing observational data. Observational scientists are using it to find important individuals, information diffusion, community structures, etc. Without an understanding of the data quality issues present in observational datasets, the results of such analyses can be misleading or biased. Bias occurs when the subjects and/or their interactions are skewed by factors such as observer interest or motivation, limited observation, or subjective interpretation. In general, bias is a lack of objectivity in data introduced by some aspect of the data collection strategy used by observational scientists. For our research purposes, we are interested in measurable bias which manifests itself as artificial skew in data such as unusual values of social network metrics and missing important edges and/or nodes. Though researchers have started examining how bias might affect these networks, a complete methodology for quantifying bias in social networks has not been developed.;In this thesis, we formally define the problem of quantifying and ranking bias in social networks and present a methodology for measuring bias in social network graphs where the underlying data is obtained through observation. We also propose a novel bias ranking algorithm that ranks bias in observed networks when compared to the ground truth network using an ensemble method which incorporates social network metrics. In order to better understand bias in the context of localized community structures, we propose a method for quantifying localized bias using graph edit distance and subgraph isomorphism with a new candidate selection scheme. Finally, we present the implementation of our methodology in a graph mining and visualization tool and test our methodology on synthetic data and the Shark Bay dolphin dataset.