Nathan Kallus: Learning Surrogate Indices from Historical A/Bs Adversarial ML for Debiased Inference

Subscribe to the channel to get notified when we release a new video.
Like the video to tell YouTube that you want more content like this on your feed.
See our website for future seminars: https://sites.google.com/view/ocis/home
Tuesday, Apr 08, 2025: Nathan Kallus (Cornell University)
- Title: Learning Surrogate Indices from Historical A/Bs: Adversarial ML for Debiased Inference on Functionals of Ill-Posed Inverses
- Discussant: Rahul Singh (Harvard University)
- Abstract: Experimentation on digital platforms often faces a dilemma: we want rapid innovation but we also want to make decisions based on long-term impact. Usually one resorts to looking at indices (i.e., scalar-valued functions) that combine multiple short-term surrogate outcomes. Constructing indices by regressing long-term metrics on short-term ones is easy with off-the-self ML but suffers bias from confounding and direct (i.e., unmediated) effects. I will discuss how to instead leverage past experiments as instrumental variables (IVs) and some surrogates as negative-control outcomes, with real-world examples from Netflix. There are two key technical challenges to surmount to make this possible. First, past experiments characterize the right surrogate index as a solution to an ill-posed system of moment equations: it does not uniquely identify an index, and approximately solving it does not translate to approximating any solution. We tackle this by developing a novel debiasing method for inference on linear functionals of solutions to ill-posed problems (as average long-term effects are such functionals of the index) and adversarial ML estimators for the solution admitting flexible hypothesis classes, such as neural nets and reproducing kernel Hilbert spaces. Second, even as we observe more past experiments, we have non-vanishing bias in estimating the moment equation implied by each one, since each experiment has a bounded size that is often just barely powered to detect effects. We tackle this by incorporating an instrument-splitting technique into our estimators, leading to a ML analogue of the classic (linear) jackknife IV estimator (JIVE) with guarantees for flexible function classes in terms of generic complexity measures.

Видео Nathan Kallus: Learning Surrogate Indices from Historical A/Bs Adversarial ML for Debiased Inference канала Online Causal Inference Seminar