Deep representation learning for continuous treatment effect estimation

Estimating the individual treatment effect (ITE) from observational datasets has important applications in domains such as personalized medicine, economics, and recommendation systems. The observational datasets often exhibit treatment-selection bias, resulting in a distribution shift among populations of samples that received different treatments. While deep representation learning has shown great promise in adjusting for covariate shifts when the treatment is a binary variable, the more practical and challenging task of handling continuous treatments (e.g., dosage of a medication) remains relatively underexplored. In this thesis, our aim is to address the associated challenges with continuous treatment. Specifically, we propose a deep model that mitigates the distribution shift through an adversarial procedure and predicts the potential outcomes using an attention mechanism. The model's objective is grounded in a theoretical upper bound on counterfactual prediction error. Our experimental evaluation on semi-synthetic datasets also demonstrates the method's empirical superiority over a range of state-of-the-art.