GraphITE: Estimating Individual Effects of Graph-structured Treatments
This addresses a domain-specific problem in causal inference for applications like drug discovery, where treatments are complex and numerous, but it is incremental as it builds on existing methods for treatment effect estimation.
The paper tackles the problem of estimating individual treatment effects when treatments are graph-structured, such as drugs, by proposing GraphITE, which uses graph neural networks and regularization to mitigate observation biases. Experiments on real-world datasets show it outperforms baselines, particularly with many treatments.
Outcome estimation of treatments for target individuals is an important foundation for decision making based on causal relations. Most existing outcome estimation methods deal with binary or multiple-choice treatments; however, in some applications, the number of treatments can be significantly large, while the treatments themselves have rich information. In this study, we considered one important instance of such cases: the outcome estimation problem of graph-structured treatments such as drugs. Owing to the large number of possible treatments, the counterfactual nature of observational data that appears in conventional treatment effect estimation becomes more of a concern for this problem. Our proposed method, GraphITE (pronounced "graphite") learns the representations of graph-structured treatments using graph neural networks while mitigating observation biases using Hilbert-Schmidt Independence Criterion regularization, which increases the independence of the representations of the targets and treatments. Experiments on two real-world datasets show that GraphITE outperforms baselines, especially in cases with a large number of treatments.