Batch Mode Active Learning for Individual Treatment Effect Estimation

Zoltan Puha*, Maurits Kaptein, Aurelie Lemmens

*Corresponding author for this work

Research output: Chapter/Conference proceedingConference proceedingAcademicpeer-review

Abstract

Field experimentation has become a well-established practice to estimate individual treatment effects. In recent years, the Active Learning (AL) literature has developed methods to optimize the design of field experiments and reduce their cost. In this paper, we propose a novel AL algorithm for individual treatment effect estimation that works in batch mode for cases where the outcomes of an intervention are not immediate. It uniquely combines Expected Model Change Maximization and Bayesian Additive Regression Trees. Our approach (B-EMCMITE) uses the predictive uncertainty around the individual treatment effects to actively sample new units for experimentation and decide which treatment they will receive. We perform extensive simulations and test our approach on semi-synthetic, real-life data. B-EMCMITE outperforms alternative approaches and substantially reduces the number of observations needed to estimate individual treatment effects compared to A/B tests.
Original languageEnglish
Title of host publicationProceedings - 20th IEEE International Conference on Data Mining Workshops, ICDMW 2020
EditorsGiuseppe Di Fatta, Victor Sheng, Alfredo Cuzzocrea, Carlo Zaniolo, Xindong Wu
PublisherIEEE Computer Society
Pages859-866
Number of pages8
ISBN (Electronic)9781728190129
DOIs
Publication statusPublished - 16 Feb 2021
Event20th IEEE International Conference on Data Mining Workshops, ICDMW 2020 - Virtual, Sorrento, Italy
Duration: 17 Nov 202020 Nov 2020

Publication series

SeriesIEEE International Conference on Data Mining Workshops, ICDMW
Volume2020-November
ISSN2375-9232

Conference

Conference20th IEEE International Conference on Data Mining Workshops, ICDMW 2020
Country/TerritoryItaly
CityVirtual, Sorrento
Period17/11/2020/11/20

Fingerprint

Dive into the research topics of 'Batch Mode Active Learning for Individual Treatment Effect Estimation'. Together they form a unique fingerprint.

Cite this