# Proxy Methods for Domain Adaptation

**Authors:** Katherine Tsai, Stephen R. Pfohl, Olawale Salaudeen, Nicole Chiou, Matt J. Kusner, Alexander D'Amour, Sanmi Koyejo, Arthur Gretton  
**arXiv:** [2403.07442](https://arxiv.org/abs/2403.07442)  
**Date:** 2024-03-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.07442v1/

## Key points

- Authors: Katherine Tsai, Stephen R. Pfohl, Olawale Salaudeen, Nicole Chiou, Matt J. Kusner, Alexander D'Amour, Sanmi Koyejo, and Arthur Gretton
- Problem: Domain adaptation under distribution shift due to changes in the distribution of an unobserved latent variable affecting covariates and labels
- Approach: Proximal causal learning technique for estimating causal effects with available proxies of unobserved confounders
- Settings considered: Concept Bottleneck (additional "concept" variable mediates relationship) and Multi-domain adaptation (training data from multiple source domains with varying distributions over latent confounder)
- Methodology: Two-stage kernel estimation approach developed to address complex distribution shifts in both settings
- Validation: Experiments demonstrating superiority over methods explicitly recovering latent confounder; small-scale experiment using chest X-ray data from MIMIC-CXR dataset for classification tasks related to radiological findings
- Results: Effective handling of distribution shifts similar to those observed in previous studies; promising proxy-based approach for real-world applications

## Summary

In their paper "Proxy Methods for Domain Adaptation," Katherine Tsai, Stephen R. Pfohl, Olawale Salaudeen, Nicole Chiou, Matt J. Kusner, Alexander D'Amour, Sanmi Koyejo, and Arthur Gretton explore the problem of domain adaptation under distribution shift caused by changes in the distribution of an unobserved latent variable that affects both covariates and labels. They introduce a novel approach to adaptation using proximal causal learning - a technique that estimates causal effects when proxies of unobserved confounders are available. This method allows for adaptation to distribution shifts without explicitly modeling or recovering latent variables. The authors consider two specific settings: Concept Bottleneck and Multi-domain adaptation. In the Concept Bottleneck setting, an additional "concept" variable mediates the relationship between covariates and labels. In the Multi-domain setting, training data from multiple source domains with varying distributions over the latent confounder is utilized. To address complex distribution shifts in both settings, they develop a two-stage kernel estimation approach. To validate their approach, they conduct experiments demonstrating its superiority over other methods that explicitly recover the latent confounder. Additionally, they perform a small-scale experiment using chest X-ray data from the MIMIC-CXR dataset for classification tasks related to radiological findings. The results showcase the effectiveness of their method in handling distribution shifts similar to those observed in previous studies. Overall,this research contributes valuable insights into domain adaptation under distribution shift scenarios and presents a promising proxy-based approach for addressing such challenges effectively in various real-world applications.