Main project: In many common fairness settings, we can’t observe counterfactuals — e.g. would a person we denied a loan pay it back if we had granted it? — and this biases training data, creating uncertainty particularly for groups historically denied loans.
Using techniques like structural equation modeling and infomax active learning, we seek to formalize whether getting labels (say, granting loans) for underrepresented groups or areas of our feature space for informational purposes could improve our model faster. This presents an explore-exploit tradeoff common in sequential decision problems, where we must balance doing what our model currently recommends with acquiring better training data.
Technically, using blavaan for Bayesian latent-variable modeling, and formulated likelihood function for our version of SEM