My research focuses on social media as a source of research questions and a source of data. No other technology has changed everyday lives as profoundly and rapidly as social media. Platforms like Twitter, Facebook, and Tiktok now connect billions of people within online spaces to exchange ideas and information, work, socialize, date, entertain themselves and even fight wars. These massive, global interconnections promote liberty, openness and the free exchange of ideas. However, due to their low barrier to entry and global reach, social platforms have become a target for social manipulation by malicious actors who aim to spread misinformation, inflame culture wars, and create polarization. My research attempts to reduce these harms and increase the benefits of interconnectedness through the synthesis of social networks and AI.

Bias in Data and AI Fairness

Our reliance on data to fuel AI raises important questions about fairness and ethics. Social data is often heterogeneous, as it comes from a population composed of subgroups with different characteristics and behaviors. A trend in aggregate data may disappear or reverse when the data is disaggregated into its constituent subgroups. This effect, known as Simpson’s paradox, often confounds models learned from data. We are developing methods to quantify biases in heterogeneous data. One approach leverages the Simpson's paradox by accounting for latent groups to learn more robust and generalizable models. We are also developing principled mathematical methods to create unbiased features for learning fair models, or use affirmative action to improve collective outcomes of interventions.

Gender Bias in Science

Despite long-term efforts to increase women’s representation in the scientific workforce, they continue to face barriers to advancement. We showed that women publish in less prestigious journals and receive fewer citations. The multifaceted gender disparities create a glass ceiling, an invisible barrier that fundamentally limits professional opportunities for even the best women scientists. The Covid-19 pandemic has only amplified existing gender disparities. To address these challenges, we are developing methods to audit gender biases in science. One recent example is a PNAS paper which identified gender disparities in the citations of members of the National Academy of Sciences. These differences, moreover, were strong enough to enable us to accurately predict the scholar’s gender from their citation networks. We have also developed a model of growing citation networks that explains the emergence of gender disparities in science.

Friendship Paradox in Social Networks

People’s observations of their friends within social networks are systematically biased. One of the better-known sources of bias is the friendship paradox, which states that people are less popular than their friends, on average. The friendship paradox is everywhere in online social networks of Twitter, Instagram and the like. However, there is more: friendship paradox also holds for any trait that is correlated with popularity. As a result, your friends (and followers) on Twitter post more messages and receive more diverse and novel information than you do. It gets more interesting ... We showed that a stronger version of the paradox holds: most of your friends are more popular than you are (not on average)! What’s more, any trait correlated with popularity will again lead to a paradox: for example, the vast majority of Twitter users will observe that most of their friends (and followers) are more active than they are, and also have more friends and followers. Strong friendship paradox leads to mind-bending phenomena, like the Majority Illusion, in which a rare trait can appear to be exceedingly popular within most social circles. This can bias perceptions of popularity online: we showed that some topics on Twitter appear to be several times more popular than they really are, i.e., many more people see their friends talking about it compared to how many people are actually talking about it. Explaining these paradoxes mathematically required us to define a new property of social networks -- transsortativity -- which measures the correlation of the popularity of a node’s neighbors. Friendship paradox explains why your friends seem to lead more exciting lives than you do. On the downside, it could fuel negative social comparisons that are detrimental to mental health and wellbeing.