How does fairness interact with efficiency when online marketplaces use profile photos?
With job market candidate Emil Palikot, Dean Karlan, and Yuan Yuan, in a new paper we combine observational data, experiments, and a calibrated model to do counterfactual analysis. Which borrower will appeal most to lenders, and how does that depend on the borrower's "type" and their "style"?
In this paper, we make use of computer vision techniques to measure attributes of profile photos on the Kiva platform and estimate the effects of these attributes on lender choices. But, this is observational data and omitted variables may be present! So we use GANs to create altered images that differ in only one dimension at a time, and estimate preferences in a survey experiment.
The experimental estimates are closely aligned with observational study estimates. We use them to study what would happen if marketplace naively promotes profiles with "style" features like smiles and closeup photos, as might occur with a rec engine incorporating image features.
We find that naively incorporating image features exacerbates inequity by borrower types. But encouraging participants to select photos with desired features - "say cheese!" - can improve efficiency and equity. https://arxiv.org/abs/2209.01235
Very cool! You may be interested in this paper from Elisa Macchi, where she manipulated images of loan applicants in Uganda to add weight--a signal of wealth.