Challenging power in data science

Context 

The high level of interest in this event (over 700 people subscribed) reflects the need to understand and challenge power in data science. Moreover, the ongoing protests against racism triggered by the killing of George Floyd in Minnesota have made it even more pertinent to consider and identify the structural power dynamics — and the means of resistance—that are recounted through lucid examples in Data Feminism. 

The overarching premise throughout  Data Feminism is to identify the imbalanced power structures underlying data science. One way to get a sense of how inequality is built into data science is to consider cases of biased technological products. For example, discriminating AI products and algorithms have led police in the US to unfairly target certain neighbourhoods with high proportions of people from ethnic minorities, regardless of the area’s crime rate.  

The way data is used can also highlight inequalities. D’Ignazio and Klein offer O’Neil’s example (elaborated in Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy) of New York’s police employing property tax delinquency data and arrest locations to predict where future crimes will take place. As a result, poorer neighbourhoods were patrolled more heavily. The very presence of a greater police scrutiny, in turn, meant increased reports of crimes and police activity, feeding into what O’Neil calls a “pernicious feedback loop.” The purpose of the data collection was to predict where and when crimes would occur, but its application proved to reinforce the social injustices already in place.  

A critical point that is made is that Data Feminism is not solely about gender or sex, but also about power structures that affect us as individuals. By taking this broader view, data feminism is intersectional. This means that our experiences are affected by diverse factors, such as gender or sex, but also other factors such as race, class, and disability. We all exist within systems of power designed with some people in mind and not others, which, as a result, maintain systemic advantages for some while reinforcing disadvantages for others.  

These systems interlock and interact with one another. For example, an individual might experience oppression because they are a woman, while at the same time experience a certain privilege because they are white. Patricia Hill Collins calls these interlocking systems ‘the matrix of domination’: this matrix upholds power for those who already hold it and oppresses those who are already marginalised.  

Data feminism, as a movement, also highlights the inadequacy of making gender or sex “categories” binary. One problem is that taking gender as binary (seeing only “men” and “women”) ignores the transgender community. Constructivist accounts of gender, which take gender as a social construct, also invite us to question the nature of such distinctions. In scrutinising the man/woman divide, it becomes apparent that there is an underlying mechanism that perpetuates a hierarchy that oppresses women. Identifying this mechanism (the talk of gender or sex being binary) is the first step towards abolishing gender inequalities. 

Three aspects of data feminism 

D’ignazio and Klein’s talk invited the audience to reflect on three aspects of data science: the production of datasets and data that is missing, the importance of data pluralism, and the role of emotion in data visualisation. 

On missing data, Mimi Ọnụọha was introduced, an artist and researcher whose work highlights the social relationships and power dynamics behind data collection. Through her work, we learn of the datasets that are not collected, information that is not released and knowledge that is not shared despite these “blank spots” existing in otherwise data-saturated spaces. The question we must ask is why? Why might some data not be collected? Ọnụọha gives four possible responses, each with real-life examples. The underlying issue highlighted with each example is the role of power and how influence is often exerted by decision makers to keep data behind closed doors which effectively maintains the status quo. 

D’Ignazio and Klein also discussed the importance of pluralism in data science: the need to include multiple perspectives (emphasising those of local and indigenous people) for comprehensive analyses. They compared the work of the San Francisco-based Anti-Eviction Mapping Project (AEMP)—whose work is grounded in community-organising and local knowledge—with the academic work of Princeton’s Eviction Lab. The Data Feminism authors explain how the Eviction Lab had approached the AEMP to work with them and their data, but conversations on how to safeguard data stalled and the Eviction Lab ultimately purchased data from a data-broker. The Eviction Lab’s decision prioritised speed and comparability of data at the expense of community knowledge and trust. This also meant decreased accuracy, as depicted by the AEMP finding three times the number of evictions in California compared to the Eviction Lab (see also p.31 of Housing justice in unequal cities).  A shared goal—tracking evictions and making them visible — is not sufficient: science is enhanced by collaboration. 

On the role of emotion, D’Ignazio and Klein spoke of the role of values in data visualisation. They mention how, for example, simply adding a URL (even a fake one) will make a graph be deemed more credible. They argue that data visualisation should be used to conjure up emotions in their interpreters. D’Ignazio and Klein argue that rather than placing importance on distance and ‘neutrality’, data scientists—and particularly data visualisers—should recognise that data is produced from a particular standpoint and that excluding emotion and feelings leads to a partial  analysis. 

Consider, for example, this visualisation of gun deaths in the US. The emotion the graph seeks to invoke mirrors how its designers believe you should feel about the data represented. D’Ignazio and Klein argue that recognising this aspect of data visualisation produces better, more holistic analyses that can, in turn, lead to policies which drive meaningful change.

Visualisation of gun deaths in the US in 2013 by Periscopic