SemiITE: Semi-supervised Individual Treatment Effect Estimation via Disagreement-Based Co-training

Qiang Huang, Jing Ma, Jundong Li, Huiyan Sun, Yi Chang

Abstract: Recent years have witnessed a surge of interests in Individual Treatment Effect (ITE) estimation, which aims to estimate the causal effect of a treatment (e.g., job training) on an outcome (e.g., employment status) for each individual (e.g., an employee). Various machine learning based methods have been proposed recently and have achieved satisfactory performance of ITE estimation from observational data. However, most of these methods overwhelmingly rely on a large amount of data with labeled treatment assignments and corresponding outcomes. Unfortunately, a significant amount of labeled observational data can be difficult to collect in real-world applications due to time and expense constraints. In this paper, we propose a Semi-supervised Individual Treatment Effect estimation (SemiITE) framework with a disagreement-based co-training style, which aims to utilize massive unlabeled data to better infer the factual and counterfactual outcomes of each instance with limited labeled data. Extensive experiments on two widely used real-world datasets validate the superiority of our SemiITE over the state-of-the-art ITE estimation models.