Fri, May 07, 2021 @ 09:00 AM - 11:00 AM
PhD Candidate: Nazanin Alipourfard
Date: May 7, 2021
Dissertation defense committee:
Kristina Lerman (chair), Ellis Horowitz, Jose-Luis Ambite, Greg Ver Steeg, Phebe Vayanos
Emergence and Mitigation of Bias in Data and Networks
The presence of bias often complicates the quantitative analysis of large-scale heterogeneous or network data. Discovering and mitigating these biases enables a more robust and generalizable analysis of data. This thesis focuses on the 1) discovery, 2) measurement and 3) mitigation of biases in heterogeneous and network data.
The first part of the thesis focuses on removing biases created by the existence of diverse classes of individuals in the population. I describe a data-driven discovery method that leverages Simpson's paradox to identify subgroups within a population whose behavior deviates significantly from the rest of the population. Next, to address the challenges of multi-dimensional heterogeneous data analysis, I propose a method that discovers latent confounders by simultaneously partitioning the data into fuzzy clusters (disaggregation) and modeling the behavior within them (regression).
The second part of this thesis is about biases in bi-populated networked data. First, I study the perception bias of individuals about the prevalence of a topic among their friends in the Twitter social network. Second, I show the existence of power-inequality in author citation networks in six different fields of study, due to which authors from one group (e.g., women) receive systematically less recognition for their work than another group (e.g., men). As the last step, I connect these two concepts (perception bias and power-inequality) in bi-populated networks and show that while these two measures are highly correlated, there are some scenarios where there is a disparity between them.
Audiences: Everyone Is Invited
Contact: Lizsl De Leon