Peeking into the Black Box: An Intersectional Perspective

• In Event: Methods to Study Diversity

Abstract

In the spirit of methodological pluralism, we investigate the compatibility of machine learning and intersectionality. Intersectional quantitative methods would require the researcher to underscore and specify the important interdependent relationship between race, gender, and class; particularly as it relates to the political experience of multiply marginalized groups. Traditionally, machine learning treats inputs and their interactions in the model building process as black boxes, making it unsuitable for specifying and understanding group based heterogeneity. However, recent advancements have made it possible to find the attributes and dependencies used by black box models when making predictions. We utilize a gradient boosting machine (GBM) to predict voting behavior for Clinton in the 2016 election. We subsequently use an investigative function, to analyze the main predictors and interactions found by the GBM. Using these algorithms, we investigate the following questions: Do black box models pick up on relationships between race and gender that look like intersectionality? If black box models don’t pick up on relationships between race and gender that look like the relationships that are described by scholars of intersectionality, then why don’t they? By investigating the predictors and interactions in these black boxes, we investigate the utility or limitations of common machine learning methods to uncover the interdependent relationship between race, gender, and class.

Authors

• Melina Much, UC Irvine

• Kyle McWagner, University of California Irvine