EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

2321. A predict-then-optimize approach for uplift modeling with continuous individual treatment effects

Invited abstract in session MC-31: Causal Machine Learning, stream Analytics.

Monday, 12:30-14:00
Room: 046 (building: 208)

Authors (first author is the speaker)

1. Simon De Vos
Faculty of Economics and Business, KU Leuven
2. Wouter Verbeke
Faculty of Economics and Business, KU Leuven

Abstract

Uplift modeling facilitates decision optimization by predicting the instance-dependent impact of treatments on specific outcomes. In a setting with binary treatments, cost-sensitive causal classification enables the classification of instances into treated or untreated groups, aiming to maximize the expected causal profit which is the core objective. We extend upon the binary treatment case and allow for continuous individual treatment effects, represented by a dose-response curve. From an application perspective, the motivation for integrating continuous treatments stems from the potential convexity of individual dose-response curves. Deciding which entities to assign what dosage is a function of individual dose-response curves, which reflect the change in positive outcome probability per assigned dosage, and the cost and benefit parameters of treatment and positive outcomes of the specific problem setting at hand. Our predict-then-optimize approach involves two steps. First, individual dose-response curves are estimated on observational data. Since historically treatments are not assigned at random but the allocation of doses depends on pre-treatment covariates, observational data may be subject to selection bias. To address this bias, we employ causal machine learning methods. Second, we formulate the individual dosage allocation as an optimization problem to maximize the expected causal profit while adhering to constraints, such as budget limitations or fairness criteria.

Keywords

Status: accepted

Back to the list of papers