Graph Learning for Uplift Modeling


George Panagopoulos

George Panagopoulos

George Panagopoulos is a postdoctoral scientist at the University of Luxembourg, researching graph neural networks and causal inference for biomedical applications. Before that, he was working on machine learning for operations research as an applied scientist in the Algorithms and Optimization lab of Amazon Transportation Services in Luxembourg. He received his PhD in machine learning from the Data Science and Mining Team of the École Polytechnique, specializing in graph learning for forecasting and combinatorial optimization. Previously, he was a research assistant and obtained his M.Sc. in computer science with a focus on digital signal processing for neural/physiological data from the University of Houston.


From precision medicine and drug discovery to recommendation systems and online marketing, causal inference using randomized experiments is the gold standard for decision-making.

However, these experiments require interventions that might be too costly, time-consuming, or simply impossible. Machine learning has provided promising solutions in predicting the conditional average treatment effect of an intervention, without actually making it [1]. In this project, we will examine the use of machine learning to facilitate a large-scale marketing campaign.

During such a campaign, promotional codes are distributed to users to increase their activity. Given a specific promotional budget, the campaign should minimize the outreach to users who will not respond, or even worse respond negatively. Hence the aim is to build a model that can predict which users will respond positively to an intervention, commonly called uplift modelling. In uplift modeling, we run an experiment on a representative subset of the users and train a model to predict the average treatment effect on the rest of the user base.

The plan is to focus on the network of interactions between samples, which can hide potential confounders that introduce bias in the experiment [2]. We will tackle uplift modeling using graph learning on a heterogeneous graph of an e-commerce system with ground truth interventions and outcomes from an actual marketing campaign. We will then compare the performance with state-of-the-art methods using one of the libraries for ML-based causal inference [3].

[1] Künzel, Sören R., et al. “Metalearners for estimating heterogeneous treatment effects using machine learning.” Proceedings of the national academy of sciences 116.10 (2019): 4156-4165.

[2] Guo, Ruocheng, Jundong Li, and Huan Liu. “Learning individual causal effects from networked observational data.” Proceedings of the 13th international conference on web search and data mining. 2020.

[3] Syrgkanis, Vasilis, et al. “Causal inference and machine learning in practice with econml and causalml: Industrial use cases at microsoft, tripadvisor, uber.” Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 2021.