Increasingly, practitioners are turning to ML to build causal models, and predictive models that perform well under distribution shifts. However, current techniques for causal inference typically rely on having access to large amounts of data, limiting their applicability to data-constrained settings. In addition, empirical evidence has shown that most predictive models are insufficiently robust with respect to shifts at test time. In this talk, I will present my work on building novel techniques addressing both of these problems.
Much of the causal literature focuses on learning accurate individual treatment effects, which can be complex and hard to estimate from small samples. However, it is often sufficient for the decision maker to have estimates of upper and lower bounds on the potential outcomes of decision alternatives to assess risks and benefits. I will show that in such cases we can improve sample efficiency by estimating simple functions that bound these outcomes instead of estimating their conditional expectations. I will present a novel algorithm that leverages these theoretical insights.
I will also talk about approaches to deal with distribution shifts using causal knowledge and auxiliary data. I will discuss how distribution shift arises when training models to predict contagious infections in the presence of asymptomatic carriers. I will present a causally-motivated regularization scheme that enables prediction of the true infection state with high accuracy even if the training data is collected under biased test administration.
Speaker Biography
Maggie Makar is a PhD student at CSAIL, MIT. While at MIT, Maggie interned at Microsoft Research, and Google Brain. Prior to MIT, Maggie worked at Brigham and Women’s Hospital, studying end-of-life care. Her work has appeared in ICML, AAAI, JSM, the Journal of the American Medical Association (JAMA), Health Affairs, and Epidemiology among others. Maggie received a B.Sc. in Math and Economics from the University of Massachusetts, Amherst.